Share

The Chronological Illusion: Aging Isn’t Linear

It is well recognized that each of us age at a different pace, exemplified by person aged 90 who has never been sick and physically appears to be decades younger (what we’ve called the Wellderly or Super Agers). But the overall assumption has been that all of us undergo a gradual, steady, linear process of aging, predictable wear and tear throughout lifespan, as assessed via the calendar. That “chronological illusion” (the dotted line below) has now been debunked. Multiple studies that used large-scale proteomics (proteins or other omics/biological metrics and imaging) and have found 3 major waves of aging that occur around ages 34, 60 and 78 years, as shown below.

Simplified redraw of Figure 1 from the review which cited 7 studies, made with ChatGPT

Subscribe now

Aging Is Asynchronous in Our Body

It also turns out our organs age at a different pace. We might not have known that without the ability to do high-throughput proteomics with AI analytics. By that I mean being able to assess up to 11,000 proteins from a tube of blood and use AI to define which proteins are tracking which organ and the immune system. Perhaps the best way to prove this asynchronous process within our body is from the world’s oldest longitudinally followed birth cohort— the UK’s National Survey of Health and Development (NSHD). All of the 1,800 participants were born in the same week in 1946. At age 63, proteomics were assessed for 6 organs and the immune system. In the Figure below, you can see there is about a 10-year spread of organ and immune system pace of aging among the cohort. Conventional, top left, refers to whole body, that is all measurable age-related proteins without breakdown by organ.

Within participants there were differences in the chronological age compared the person’s organ age. In the example below, the heart shows accelerating aging. Multiple studies of organ clocks have correlated fast aging of an organ with a disease indexed to that organ, such as heart to heart failure or brain to Alzheimer’s disease.

This pattern of asynchronous organ-specific aging has been noted in several proteomic studies, and has been shown to correlate with lifestyle factors (such as alcohol intake and liver age) and as I previously reviewed in Ground Truths hormone replacement therapy and the immune system age. To give you a sense of the breadth of what organ clocks tracked in the NSHD cohort, from the paper: “The organ clocks were adversely affected by overweight in adolescence, lack of education, smoking and various types of adversity across the life course (such as parental divorce or lack of social support). In contrast, there were several healthspan-protective factors including normal birthweight, higher childhood socioeconomic status, sustained physical activity and moderate alcohol intake.”

Cell Clocks and the Master Regulators of Aging

Discovering organ clocks can be considered a breakthrough in their own right. But it goes deeper—to cell clocks. Frankly, I would not have expected we could track aging for different cell types, like brain astrocytes or microglia, from the proteins appearing in the blood. But that is what is now possible. This work was published just 2 weeks before our review and included in it. Over 40 cell types were assessed in over 60,000 individuals with over 7,000 proteins biological aging. About 1 in 4 participants had accelerated aging in 1 cell type. A compelling pattern was seen for many conditions, as illustrated below for brain astrocytes and the risk of Alzheimer’s disease (AD). People with accelerated aging (“old”) of these cells had a 12.6 fold higher rate of developing AD compared with slow (“young”) aging of astrocytes.

In a similar way to assess risk of young or old astrocytes, but now broken down by the known APOE4 risk allele, with 1 copy in the middle graph and 2 copies at the far right. You can see the minimal risk in the people without E4 (at left below, 2 copies of E3), risk of AD increased with old astrocytes in those with 1 E4 allele copy, but dramatically increased risk (nearly 40-fold) for old astrocytes vs young in people with 2 copies of the E4 allele. Noteworthy, the age of astrocytes predicted risk of Alzheimer’s independent of APOE4 status. It adds to p-tau217 and other neuroinflammation markers (p-tau181, NfL, GFAP) as a way to more precisely determine risk of Alzheimer’s disease. Put another way, simple people with APOE 4 are not at high risk for Alzheimer’s, but we didn’t have a way to know that before.

A consistent patten between organ and cell clock studies is that the more evidence for accelerated aging, reflected by more organs or cells, the more increased mortality during follow-up (or reduced survival), seen side-by side below for 2 cohorts (one showing mortality and the other survival).

The same pattern for aged cell types in the UK Biobank and relationship to survival is seen in the Figure below

The other key consistency is the prominence of the relationship of survival seen with 2 organ and cell types: the brain (neurons) and immune system (immune cells). When these exhibit normal paced aging, survival over 15-17 years approaches 100%. Accordingly, the brain and immune system have the appearance of master regulators of the aging process. They seem to be acting as systemic gatekeepers, supporting resilience despite extreme aging of other organs or cells. I’ve previously emphasized the interdependence of the brain and immune system. The brain can be viewed as “mission control” for the immune system.

It isn’t as if these organs and cell types are the only ones that are informative. That couldn’t be further from what has been seen. The aging of lung cells, smoking status, and the incidence of lung cancer, as shown below, is just another of many examples.

Impact on Routine Clinical Practice

We’re not there yet to implement organ and cell clocks, but not far away. I’ve had my organ clocks assessed by one of the companies in beta phase testing. Beyond more than 10 studies cited in the review, much more validation of organ and cell aging clocks is currently underway. This month we are launching a randomized controlled trial in participants with high-risk of Alzheimer’s disease testing aggressive lifestyle improvement with 24/7 coaching to prevent the disease. All individuals will have p-tau217 and brain clocks assessed at baseline and follow-up. The main endpoint is reduction in p-tau217 after one year, which has been shown to be particularly mutable in small intervention studies with exercise (previously reviewed in Ground Truths). All of the data thus far suggest that organ clocks are eminently mutable too, influenced by lifestyle factors, medications, and many other perturbations as reviewed above.

Let’s take a step back on how this field of aging has evolved. The concept of metrics to assess biological aging got its roots more than 15 years ago with the seminal work by Steve Horvath and others assessing DNA methylation (epigenetic) clocks. Here’s my conversation about epigenetic clocks with Steve Horvath. That first generation was built on with other information, be it from plasma proteins or clinical data, to yield more refined molecular clocks. But only in the past few years, have we seen the field move from assesing human aging more holistically to getting specificity at the organ and cell levels. The Table below from our review summarizes what can be considered 6 generations of biological clocks.

This progress has implications for the practice of medicine and the potential for achieving reversal of aging. Detecting accelerated aging of an organ could enable a new path to prevention. Take a person with risk factors for heart disease who exhibits accelerated aging of his artery clock and/or blood vessel cell clock; there’s potential for a new dimension of risk to be ascertained. There is no shortage of means to help prevent atherosclerotic heart disease for such individuals. Perhaps more striking is a new window into the aging of our immune system, for which we have no way to assess in the clinic today, but now could be quantified at both the immune system and immune cell levels. That could provide the advance warning for the need to amp up the immune response, be it by lifestyle factors, vaccines or medications. And don’t forget that the same clocks can be used to assess the impact of the preventive strategy or intervention. For the 2-stage prevention of neurodegenerative diseases, the data from clocks would be added to other layers of data to assess concordance and directionality (Figure below, abbreviations PRS-polygenic risk score, EHR-electronic health record, MCI-mild cognitive impairment).

I won’t go through all the organs (or 40 cell types) but I hope this gives you the sense of enabling precision prevention.

There is considerable investment and excitement about partial cellular reprogramming (also known as “rejuvenation”) and other biotech interventions to slow the human aging process. Last month cellular reprogramming (with 3 of the 4 Yamanaka transcription factors) was initiated in a pilot trial of patients with advanced eye diseases. While that may show aging of retinal cells is slowed, it still would not demonstrate our ability to slow or reverse human aging. But comprehensive organ and cell clock assessment may ultimately provide surrogate endpoints to support efficacy and pre-empt many years of clinical trial follow-up. That’s much further away than using clocks in the clinic, but a realistic possibility if validation for such an approach is achieved.

We have a problem right now with some aging clocks. They (mostly methylation, epigenetic) are being sold commercially direct-to-consumer without standardization or reliability, and at considerable expense. I would suggest not buying any of these tests. The tests with the most potential are at the organ and cellular level, and none are commercial at this point. That will change in the next year, but will still require attention to adequate validation and standardization, and the need to bring the cost way down, which is possible when done at scale. We’ve seen that at the UK Biobank, which was able to do 6,000 plasma proteins (O-link)for approximately 500,000 participants at ~$50 a pop (original charge for O-link and Somascan was >$900 for each individual).

Like so many other biomedical advances I have recently reviewed, we’re seeing the product of very large datasets and AI analytics. That was the case for the 14-protein blood test predicting and preventing lung cancer, the breakthrough blood test for assessing the tumor microenvironment, and use of imaging the thymus for immune system health Add organ and cell clocks to that quickly growing list.

I hope I haven’t lost you along the way, as this post was fully intended to be practical, a translation of our publication with high density information and the usual jargon that makes it hard to ingest, especially for folks not grounded in this stuff. For those who have been following Ground Truths for the past few years, you will recall how excited I was about the discovery of organ clocks back in late 2023. The field has gone forward at torrent pace since then, with the added dimension of cellular clocks.

The transition from calendar to clocks in medicine will take longer than it should. It requires a structural shift in how we define age. It doesn’t mean we abandon
chronological age, but counting candles is fixed and will be transcended. Biological age is now measurable at a granular level and increasingly shown to be modifiable. I believe this represents the most important advance we have yet seen in the science of human aging. It paves the way for a great inflection for promoting health and preventing diseases in the future.

******************************

A big thanks to Ground Truths subscribers from every US state and 212 countries. Your subscription to these free essays and podcasts makes my work in putting them together worthwhile. If you’re not a subscriber, please join!

If you found this interesting PLEASE share it!

Share Ground Truths

Paid subscriptions are voluntary and all proceeds from them go to support Scripps Research. They do allow for posting comments and questions, which I do my best to respond to. Please don’t hesitate to post comments and give me feedback. Let me know topics that you would like to see covered.

Leave a comment

Many thanks to those who have contributed—they have greatly helped fund our summer internship programs for the past two years. It enabled us to accept and support a record number of 51 summer interns here with us now! These are high school, college and medical students selected from thousands of applicants. We couldn’t do this expanded program without the funds coming in through Ground Truths.

Recently, Professor Aaron Newman at Stanford and colleagues from many US leading centers published a landmark paper in Nature. It may ultimately be considered one of the most important leaps in cancer diagnostics in decades.

This was a massive undertaking involving 10 million single-cell RNAseq in 10 types of cancer to define 9 TME types by spatial omics which they called “spatial ecotypes”. It was AI-enabled by machine learning (Spatial Ecotyper from biopsy samples) and extrapolating that to knowledge to deep learning AI (Liquid Ecotyper from blood samples). The whole project took several years.

Here’s an infographic summary I made with the help of NotebookLM along with the Figures that follow.

Some of the key points from our conversation:

—TME is the soil in which the cancer cells live. It strongly influences whether the cancer cells will thrive or not, spread, respond to therapy, and determine the fate pf pre-cancerous lesion (as seen in recent important paper in pancreatic cancer evolution and colon cancer development).

—Ecotypes, a term coined in 1922, to get at the local environment of tissue and cancer, was repurposed in this new project. Good metaphor: neighborhoods. And emphasis on location , location, location like real estate.

—In the clinic today, the only window to TME is via a biopsy. But decisions for treatment are not based on TME, they are based on biomarkers like tumor mutation burden (TMB), PD-L1 levels , or volume of tumor DNA by liquid biopsy.

Share Ground Truths

—TME is dynamic and changes over time so a single snapshot won’t cut it.

—Spatial 3D maps of TME from 10 million cells, 10 cancer types led to 9 spatial ecotypes (with the Spatial Ecotyper machine learning). Validation across 3D grids (such as Visium) and multiple cohorts. The cell location base within TME is striking.

—These 9 neighborhoods are conserved and very different. SE4-immune system is asleep, hypoxic; SE 7/8 are “war zones” rich with immune cells, hot tumor; SE 5 exhibits intense immunosuppression, hostile vs immune system, portends poor survival, SE9 deep in core, driving angiogenesis (new blood vessels)

—Now rich information from these 9 SEs needs to be extrapolated to a holistic blood test of TME. It turns out an extension of tapping into methylation as used by some liquid biopsy cell-free DNA tests works very well, providing cell identity from the DNA fragments leaked into the blood. To get to TME spatial biology in a tube of blood. Akin to wastewater surveillance form municipals for detection of infectious disease pathogen sequence variants and circulating levels. A triumph of AI analytics made this possible!

—The Liquid Ecotyper was validated in 78 patients with malignant melanoma pre-treatment from Yale University. SE 7/8 benefited from immune checkpoint therapy; SE4 resisted treatment, shorter survival. Matched with tissue biopsy work.

—Work beyond publication ongoing in many types of lung cancer, bladder cancer, mesothelioma, and response to other therapies like CAR T and bispecific antibodies. This is critical since we have slew of different cancer immunotherapies (see my pyramid Ground Truths review) and we’re not sure who, when, how to use them to drive optimal cancer outcomes in patients. May also help predict other types of therapy (not just immunotherapy) for cancer.

—The blood test can performed serially which is a major advance compared with how we are currently “flying blind” and misled by imaging.

—Started a company Liquid Cell Dx to make the TME blood test commercially available. May be available sooner in the academic centers in the current study.

—The use of SEs may extend beyond cancer!

A big thanks to Ground Truths subscribers from every US state and 212 countries. Your subscription to these free essays and podcasts makes my work in putting them together worthwhile. If you’re not a subscriber, please join!

If you found this interesting PLEASE share it!

Share Ground Truths

Paid subscriptions are voluntary and all proceeds from them go to support Scripps Research. They do allow for posting comments and questions, which I do my best to respond to. Please don’t hesitate to post comments and give me feedback. Let me know topics that you would like to see covered.

Leave a comment

Many thanks to those who have contributed—they have greatly helped fund our summer internship programs for the past two years. It enabled us to accept and support a record number of 51 summer interns coming in 2026! These are high school, college and medical students selected from thousands of applicants. We couldn’t do this expanded program without the funds coming in through Ground Truths.

Thank you , , , , , and >400 others for tuning into my live video with ! Join me for my next live video in the app.

Get more from Eric Topol in the Substack app

Available for iOS and Android

Get the app

“I may be the expert in cancer, but you are the expert in you.”—Elisa Port, MD

Dr. Elisa Port is Professor of Surgery, Icahn School of Medicine at Mount Sinai; Chief of Breast Surgery for the Mount Sinai Health System; Director of the Dubin Breast Center; and Director of the Center of Excellence for Breast Cancer at the Mount Sinai Tisch Cancer Center, a NCI-Designated Comprehensive Cancer Center. She is remarkably well qualified to provide guidance for prevention and treatment of breast cancer, and runs a model program at Mount Sinai in New York (←we’ll get into why I say that).

She has just published a terrific book that centers around the concept of informed optimism. Many of her patient stories are integrated.

We tried to cover the waterfront during our conversation including:

—4 Ws for reduction of risk: weight, walking, wine, and working out (especially in Nature, “green exercise”)

—Only 10-12% of women will ever develop breast cancer during their lives

—Role of adipocytes (fat cells), aka estrogen and inflammation factories

—Misinformation (underlying the need for this book)

—Pervasive myths of magic supplements, with safety issues of unregulated intravenous drips and others when analyzed turned out to be crushed up, dried house plants

—Total body MRI: mixed results from her patients

—Breast cancer on the rise in young women. Dr. Port relates about recent patients diagnosed in their 20s. Some without BRCA or other pathogenic mutation. Why?

—Mammograms in non-dense breasts pick up 80-90% of breast cancer. It’s the 40-50% women with dense breasts, which the scan is trying to find a polar bear in a snowstorm, is a problem. The problem with false-positives.

—Use of AI with mammography for increased initial detection, prevention, detect high-risk in women with a “normal” mammogram, and risk of heart disease (see my recent Ground Truths and piece in The Lancet on this important topic)

—Mount Sinai provides Transpara AI (the one validated by the big Sweden trial of >100,000 women) to all women getting mammograms FOR FREE. That is the best validated AI tool in medicine today!

—How a mammogram saved a patient’s life, not about breast cancer!

—Hormone replacement and birth control pills. How to assess benefit and risk in women young and older? Differences in the hormone risks.

—Tamoxifen for primary prevention in high-risk people. Reduction of breast cancer in 50% or more in these women. In some subgroups reduction as high as 80%!

—GLP-1 drugs may reduce breast cancer risk and this may not just be related to weight loss

—Lack of utilization of Genetic testing (it could be your father). I disagreed with respect to use of polygenic risk scores, as I wrote about in SUPER AGERS

—The lack of our ability to assess the immune system in the clinic and how that could change the face of prevention and very early detection of breast cancer

An infographic summary I made on Dr. Port’s book with the help of NotebookLM

****************************************

Thank you , , , , , and 450+ others for tuning into my live video with ! Join me for my next live video in the app.

A big thanks to all the Ground Truths subscribers from every US state and 212 countries. Your subscription to these free essays and podcasts makes my work in putting them together worthwhile. If you’re not a subscriber, please join!

If you found this interesting PLEASE share it!

Share Ground Truths

Paid subscriptions are voluntary and all proceeds from them go to support Scripps Research. They do allow for posting comments and questions, which I do my best to respond to. Please don’t hesitate to post comments and give me feedback. Let me know topics that you would like to see covered.

Leave a comment

Many thanks to those who have contributed—they have greatly helped fund our summer internship programs for the past two years. It enabled us to accept and support a record number of 51 summer interns coming in 2026! These are high school, college and medical students selected from thousands of applicants. We couldn’t do this expanded program without the funds coming in through Ground Truths.

Get more from Eric Topol in the Substack app

Available for iOS and Android

Get the app

Note the Towards in the title of the 2 papers:

Summary of MIRA

This was designed to be embedded in a health system EHR to provide reasoning and action steps. There were 2 agents, the patient and the AI physician (MIRA). MIRA queried the patient’s history, the physical exam results, and could order labs, blood cultures, scans, medications, procedures, surgery, and triage for hospital admission. This was done in 500 emergency department established real cases with the MIRA results directly compared to 4 board-certified (BC) physicians, and also to a hybrid group of 2 BC physicians and 2 residents. (I won’t review the results of the hybrid group further since their performance in all tasks was lower than the 4 BC physicians.) It simulated the sequential way a patient’s data would be interrogated and processed. MIRA was enabled with 11 different tools and choices from >85,000 action options, operating in a standards-compliant framework for multi-step reasoning (using FHIR, ICD-10, RxNORM, ATC, LOINC, and SNOWMED-CT). The system was built on OpenAI’s GPT-4o.

The overall diagnostic accuracy for MIRA was 87.8% compared with BC physicians at 78.1%. That increase was especially notable for specific diagnoses like pancreatitis (95.2% vs 78.6%), and appendicitis (100% for MIRA, 88% for BC physicians). While MIRA ordered more blood tests (51% vs 28%), resource consumption was countered this by ordering substantially less scans. For therapy, MIRA surpassed the BC physicians 53.5% vs 38.3% for correctly ordering procedures such as laparoscopic appendectomy or cholecystectomy (Figure below). Other advantages for MIRA therapy included better IV fluid management and analgesic adherence to guidelines, and an overall 35% increased alignment of clinical guidelines compared with the BC physicians. Of 468 medications ordered by MIRA, 99.8% were correct for indication and safety (such as allergy, interactions, and kidney dosing). MIRA triaged more cases than physicians for hospital admission, which reflects lack of being economically driven.

Because the design of the system would allow leak of the case data to the AI, considerable effort was made to avoid premature information flow. That worked well, with 0 of 933 cases exhibiting any leak. 880 adversarial prompts were tested and the system held up well, as it did for stress-testing attempts at hacking, medico-legal threats, and other patient agent trickery. MIRA also assessed multiple patient perturbations including high anxiety, non-English speaking, paranoia, and diagnostic denial, without affecting its performance.

Summary of AMIE

This system had a very different design, with its focus on longitudinal assessment of outpatients, with the primary goal of developing first-rate management plans. Like MIRA, there were 2 agents used. The Dialogue Agent was conversational, interacting with the patient, representing fast, System 1 thinking (À la Danny Kahneman, using Gemini 1.5 Flash), and asynchronous to the Mx management agent, System 2, slow thinking that used long, context processing (even though it was quick). AMIE assessed 100 patients with 3 visits (each separated by ~2 days) spanning 5 different specialties. The results were compared with 21 BC primary care physicians. A noteworthy feature was the Ensemble Refinement which took 4 different treatment plans developed and came up with a consensus, mimicking a real medical treatment board, as is typically seen with cancer management. The massive >600 clinical guidelines were fully tokenized (not just parts of them) to provide the grounding and citations for management. This ensemble only took about 80 seconds to produce. Like MIRA, this was all text based, which was defended as necessary in AIME for the intent to maintain blinding. 30 physicians rated the performance of AIME outputs vs the 21 BC PCPs. Unlike MIRA, patient actors were used. The performance wheel below shows the differences for AIME vs the PCPs for 6 metrics, all leaning to AIME for being slightly or substantially better.

Overall, for management reasoning, AIME was non-inferior to the 21 PCPs. By the 3rd outpatient visit, the rating of AMIE’s management plan was 98% vs 81% of the PCPs. Preciseness of treatment was 95% vs 67%. Expert guidelines alignment was 100% vs 86%, respectively. So while the overall management was non-inferior, there were several ratings that favored AIME’s management.

A new benchmark for medication management was developed, called RxQA, built on 600 questions to board-certified pharmacists and the content from the UK and US formularies. AIME’s medication management outperfomed the PCPs for the correct medication, dose, duration, side-effects, and follow-up evaluation. This was found with the more difficult cases even when the PCPs were tested open-book (58% vs 48% in favor of AI).

Share Ground Truths

The Implications

The new studies raise the level of capabilities for medical AI from what has been previously studied, which were relatively narrow, mainly for support in making a diagnosis or answering questions on a medical exam or from a patient. MIRA took on autonomous agent end-to-end assessment and action plans for each patient presenting to an emergency department. AMIE was geared to provide longitudinal assessment across 3 outpatient visits. Both used only 2 agents, one interacting with the patient and the other to do the AI work.

There are many reasons to think these are preliminary findings that do not reflect the real world of medicine. Both MIRA and AMIE are text-only AI, meaning all the other things that are part of medicine, from the patient’s non-verbal communication and tone of voice to the review of actual medical images, were not included. The cases used in both studies were “clean,” complete data from established datasets. MIRA interactions were capped at 20 conversation turns. AIME used patient actors. There was nothing done that truly represents the practice of medicine, which is typically characterized by incomplete and conflicting data. The 3 outpatients visits in AIME were set up 2 days apart, hardly simulating how hard it is to get appointments with doctors! Likewise, only 5 specialties of medicine were addressed.

But there were some findings that can be viewed as filling in gaps of current medical care. Improving diagnostic accuracy in MIRA was clear, although the comparator group was very small, with only 4 board-certified physicians. Getting 100% correct diagnosis and ordering a laparoscopic procedure for appendicitis is impressive, as was the accuracy of the medications selected. The breadth of conditions in MIRA were limited (n=8) even though there was an array of diagnostic tests, procedures and surgeries as part of the management option mix. The lack of economic incentives in the practice of MIRA management is desirable and even resulted in less expansive scan tests ordered.

For AMIE, a big advantage was the longitudinal context across 3 clinic visits. Unlike real patients today, the Patient Dialogue Agent didn’t have to fill out forms. There was memory and efficiency (not what we see in US healthcare). That agent was set up to be empathetic and attentive which showed up in the patient preferences, a pretty big gradient favoring the AI.

Both of these medical AI models were remarkably precise with respect to sticking to guidelines or providing specific plans. For example, the PCP in the AMIE study wrote “give an antibiotic whereas the AI prescribed amoxicillin 500 mg orally, 3 tablets daily for 7 days, and wrote to check for a penicillin allergy.

But here is the rub. Guidelines in medicine are meant to apply to the vast majority of patients, without specific regards for needs, fears, prior patient experiences, cost, and many other factors. Much of present day guidelines are provided by “experts,” that is they represent opinions that are eminence-based, not evidence-based medicine. So this very tight, superior alignment of the AI with guidelines is not necessarily a good thing, with the claims of great “precision.” Frankly, over-adherence to guidelines may presage the loss of the art of medicine, not taking in the human factor of each patient, and the human-to-human bond that would be the foundation of the patient-doctor relationship.

Where is this headed? The large language models (LLMs) will keep getting better. In fact, the ones used in these 2 reports are already obsolete. In the agentic AI era, there could literally be hundreds of specialized agents, such as one for labs, one for scans, one for sensors, one for environmental exposures, one for genetics/genomics, and so on. Today, with these new papers, we are only seeing the rudimentary use of agents.

You can think of MIRA and AMIE as a major step forward within the constraints of a simulation, not real medicine. But the improvements in AI’s capabilities are coming fast, and it would not be surprising to see some of the benefits here extended to the actual practice of medicine. To prove that, we ideally will need randomized trials of 3 strategies to assess outcomes: (1) end-to-end medical AI ; (2) human clinicians only; and (3) combining both. That won’t happen soon since are still on a steep slope of model improvement and such a large trial would not be easy to get funded, no less executed. In the meantime, we’ve got new evidence for potential ways that generative AI will improve medical communication, diagnosis, and management.

NB This post was written by me, no AI. I have no COI related to the content of the post.

A big thanks to Ground Truths subscribers from every US state and 212 countries. Your subscription to these free essays and podcasts makes my work in putting them together worthwhile. If you’re not a subscriber, please join!

If you found this interesting PLEASE share it!

Share Ground Truths

Paid subscriptions are voluntary and all proceeds from them go to support Scripps Research. They do allow for posting comments and questions, which I do my best to respond to. Please don’t hesitate to post comments and give me feedback. Let me know topics that you would like to see covered.

Leave a comment

Many thanks to those who have contributed—they have greatly helped fund our summer internship programs for the past two years. It enabled us to accept and support a record number of 51 summer interns coming in 2026! These are high school, college and medical students selected from thousands of applicants. We couldn’t do this expanded program without the funds coming in through Ground Truths.

In this post, I’ll take you through the findings and the ramifications that it has for preventing cancer in the future. It’s a massive amount of work, so I will not get into the abundant technical details, but instead hit the high points.

Subscribe now

Let’s first go back to the CANTOS trial that set this up.

Background on an Old Randomized Trial

Back in 2017, there was publication of CANTOS, a large randomized clinical trial that assessed an anti-inflammatory medication (canakinumab, an interleukin-1β antibody) in more than 10,000 participants with prior heart attack for reduction of subsequent cardiovascular events. While the results were not compelling for this cardiovascular indication (small benefit and a risk of fatal infections), there was an unexpected outcome of reduction of lung cancer and fatal lung cancer during 5 year follow-up, both with a dose-response (Figure below for lung cancer incidence). At the highest dose of the drug, there was nearly an 80% reduction of fatal lung cancer. But the number needed to treat (NNT) with the antibody to prevent 1 person’s lung cancer was >1,000, which reflects there wasn’t a way to know who would benefit from the drug. Wouldn’t it be great if we could understand this better and identify whose lung cancer could be prevented with the interleukin-1β antibody?

Share Ground Truths

The New Study

Probing the UK Biobank

The first step was to do high-throughput plasma proteomics in over 48,000 UK Biobank participants (schematic below). This cohort had nothing to do with the CANTOS trial.

This involved assessing nearly 3,000 different proteins from the blood (O-link, ThermoFisher) and using machine learning of the data along with age, history of smoking and lung disease, to find the proteins that predicted lung cancer. There were 14 proteins that fulfilled this objective. They had 4 major categories of function indicative of deep lung cell stress: inflammation, lung surfactant production, epithelial cell secretion, and matrix remodeling (more on the latter 2 categories later). The proteins were present on average 5.6 years before UK Biobank participants who developed lung cancer. These 14-proteins (Figure below with risk of lung cancer) were assessed in 8 different cohorts with either the same (O-link) high-throughput proteomic platform or with Somalogic. The proteins were also confirmed in a Taiwanese cohort (comprised of 81% women, 62% adenocarcinomas) of whom 93% never smoked. Prediction of lung cancer incorporating the 14-proteins outperformed 2 prior lung cancer models based on demographics and smoking history (known as LCRAT and LLVP3). The 14-protein signature was also found in some people with idiopathic pulmonary fibrosis and in chronic obstructive pulmonary disease.

Replication of the 14-protein signature in 8 different cohorts (known by their acronyms like ARIC, EPIC, CKB, etc.)

Probing the Biology

There were important results of the TRACERx clinical study (Tracking Cancer Evolution through therapy [Rx]) that had to be folded in to help identify the source of the 14-proteins. That study tracked lung cancer evolution and post-surgical results. There was no correlation of the 14-proteins with more advanced lung cancer (compared with early cancers) and no reduction of the protein signature after resection of the lung tumor. So, surprisingly the 14-proteins were not coming directly from the tumor cells. What was the source of these proteins?

The next steps were to introduce the EGFR mutation to 4 different types of lung epithelial cells (basal, club, neuroendocrine, and AT2) in the mouse model to develop lung adenocarcinomas, and subsequently to test the impact of air pollution (particulate matter exposure), the EGFR mutation, and the combination. Single-nucleus sequencing of ~37,000 cells helped to get at the gene expression for the epithelial progression to cancer. As these different types of lung epithelial cells converged to the pre-cancerous KAC state (stands for KRT8+ alveolar intermediate cells, K for keratin). Notably, the 14-proteins were not secreted; these cells had lost their identity with malignant transformation. The source turned out to be healthy, wild-type (cells without the mutation) bystander cells sensing stress from the precancerous process, the formation of the KAC state.

Particulate matter activated macrophages to release interleukin-1β to drive production of the 14-proteins. The combination of interleukin-1β from PM exposure and the KRAS mutation took this 14-protein production to the next level. This was confirmed in human lung organoids with exposure to interleukin-1β, incriminating this factor as key to pushing dormant mutant cells into cancer and driving the 14-protein release.

Circling Back to the CANTOS Intervention Trial

With proof of the pivotal role of interleukin-1β, it was time to go back to the CANTOS trial participants and see how the antibody to it correlated with the development of lung cancer. Within 2,325 participants in that trial, there was a more than doubling of risk of lung cancer in this who had the 14-protein signature (high-risk, top graph below). And the high-risk group were the people who drive marked benefit from the interleukin-1β antibody (camakinumab, graph below, bottom) with an approximate halving of risk of lung cancer. The number needed to treat to prevent one case of lung cancer partitioned by the 14-proteins went from 1500 in the low-risk group to 50 in the high-risk group.

Here is a diagram to pull all of the above together (made with help from Gemini)

Share

The Clinical Significance of this Work

We don’t have any protein biomarkers that are validated for primary prevention of cancer, although there are some used as markers to help in early detection like CA-125 (ovarian cancer), CA19-9 (pancreatic cancer) and CEA (carcinoembryonic antigen, gastrointestinal cancer). The new study takes us to an unprecedented position for identifying and thoroughly validating a 14-protein signature for lung cancer.

This work took advantage of new methods that have only become available in recent years. At the top of that was the use of high-throughput proteomics to assay ~3,000 plasma proteins and use of machine learning to nail the 14 significantly linked to lung cancer. That is akin to many recent studies that started with 6,000 to 11,000 proteins in the blood and used AI to partition these as organ clocks and cell clocks, tracking their pace of aging status. this comes down to pinpointing the small number of proteins in the blood that correlate with a clinical metric or outcome. In the current study, showing these same 14 proteins found in the UK Biobank were predictive of lung cancer in 8 additional cohorts and in non-smokers strongly reinforced this as a seminal finding.

But this body of work went on to dissect the mechanism of the 14 proteins and what accounted for their production, through 4 different types of lung epithelial cells, laborious mouse model work with knock-in EGFR mutations, and extending into the known environmental hit of particulate matter from air pollution, strongly associated with lung cancer. We would have assumed the 14-proteins were coming from the cancer cells, but they weren’t. They were, challenging dogma, an indirect siren emitted from healthy cells, sensing the stress in cells transitioning to a precancerous state. Plus the confirmation of a double hit—particulate matter and an EGFR mutation—that led to the highest production of the 14 proteins. Imagine if we could identify a group of proteins that were predictive of each cancer 5+ years in advance and understand precisely the basis of the protein cluster! That’s where this is ultimately headed. And we’re just talking about the biomarker here.

The other dimension of this landmark study stemmed from serendipity. CANTOS was a trial directed at reducing heart disease, a secondary prevention for people who already had a heart attack. But the researchers backed into the finding of reduced lung cancer and fatal lung cancer with the innate immunity intervention of blocking interleukin-1β. While interesting, there was no way to practically take that to the clinic with a number needed to treat (NNT) exceeding 1,000 (to prevent 1 lung cancer) and the hazard of the antibody for serious infections. But CANTOS took the 14-protein biomarker story to a new level, by testing whether the antibody would work in the high risk group. And it did, lowering the NNT to 50 (figure below, made with Claude) in the 20% of participants who had the biomarkers. That’s a remarkable low number of people needed to treat to prevent one deadly cancer.

Now this, of course, is retrospective. We don’t yet have a clinical trial to prove that lung cancer can be reduced by 50% with the 14-protein biomarkers and the antibody treatment. That is the next step along the way. If it is proven prospectively, that would cinch a “prevmed", what I call a prevention medicine, here for preventing cancer. But if it weren’t for CANTOS the hunt for the biomarker would not have likely proceeded, no less identifying a highly promising interception, a primary prevention for the most common form of cancer diagnosed globally. And if we didn’t have the UK Biobank resource with almost 50,000 participants who had high-throughput proteomics and >17-year follow up for outcomes, the 14-protein biomarkers would not have been found. Just think that now all 550,000 UK Biobank participants have had O-link plasma proteomics, greatly enriching this resource for finding biomarkers and candidate prevmeds for many other types of cancer. Both the large CANTOS trial of >10,000 participants and the UK Biobank played an instrumental role in enabling the new study, no less many other cohorts with proteomics and outcomes for independent replication. All of this highlights how critical it is to have such clinical resources to make such pivotal discoveries that can impact medical practice and the lives of so many patients whose cancer could be prevented.

The new proteomic based work that elevated the potential for preventing cancer is complementary to the vaccine approach that revs up the immune system in a person at high-risk. These 2 primary prevention strategies could be combined, and would have even greater effect than either along in high-risk people.

What we have in this new study and context is the beginning of a roadmap, a template, to achieve primary prevention of a common and all too frequently lethal cancer. Once we can accurately predict cancer 5+ years ahead, the era of primary prevention will go into high gear.

NB This post was written by me, no AI. As noted, 3 images were made with the help of AI. I have no COI related to the content of the post.

A big thanks to Ground Truths subscribers from every US state and 212 countries. Your subscription to these free essays and podcasts makes my work in putting them together worthwhile. If you’re not a subscriber, please join!

If you found this interesting PLEASE share it!

Share Ground Truths

Paid subscriptions are voluntary and all proceeds from them go to support Scripps Research. They do allow for posting comments and questions, which I do my best to respond to. Please don’t hesitate to post comments and give me feedback. Let me know topics that you would like to see covered.

Leave a comment

Many thanks to those who have contributed—they have greatly helped fund our summer internship programs for the past two years. It enabled us to accept and support a record number of 51 summer interns coming in 2026! These are high school, college and medical students selected from thousands of applicants. We couldn’t do this expanded program without the funds coming in through Ground Truths.

Helen Pearson, PhD, is an award-winning biomedical journalist at Nature, named European Science Journalist of the Year in 2025. She teaches science writing at University College London. BEYOND BELIEF is her second, just published book. THE LIFE PROJECT was her first.

Subscribe now

The points we covered include:

—The evolution of evidence over 8 decades, moving from rationalism, expert opinion, hunches (eminence-based) to empiricism, evidence-based

—The enormous devastating adverse impact of Dr. Spock’s Baby and Child Care book from a single sentence change, leading to an epidemic of sudden infant death syndrome. And how it ultimately got corrected.

—The term evidence-based medicine was coined in 1991 at McMaster University

—Iain Chalmers, Archie Cochrane and the Cochrane Collaboration

“The scandalous failure of science to cumulate evidence scientifically”—Iain Chalmers

—Systematic reviews

—Natural experiments (such as with Shingles)

—Resistance to change

“Challenging a doctor’s conventional way of practice is like hitting him in the intellectual testicles”—Drummond Rennie, past JAMA editor

—Evidence-based practice (and not) in the Covid pandemic

—Recent examples of challenging dogma (kindly stones, thymus involution)

—AI and misinformation, false evidence

Thank you , , , , , and many others for tuning into my live video with ! Join me for my next live video in the app.

A big thanks to Ground Truths subscribers from every US state and 212 countries. Your subscription to these free essays and podcasts makes my work in putting them together worthwhile. If you’re not a subscriber, please join!

If you found this interesting PLEASE share it!

Share Ground Truths

Paid subscriptions are voluntary and all proceeds from them go to support Scripps Research. They do allow for posting comments and questions, which I do my best to respond to. Please don’t hesitate to post comments and give me feedback. Let me know topics that you would like to see covered.

Leave a comment

Many thanks to those who have contributed—they have greatly helped fund our summer internship programs for the past two years. It enabled us to accept and support a record number of 51 summer interns coming in 2026! These are high school, college and medical students selected from thousands of applicants. We couldn’t do this expanded program without the funds coming in throughGround Truths.

Get more from Eric Topol in the Substack app

Available for iOS and Android

Get the app

Subscribe now

But now we’re seeing an exponential rise in new ways, clinical trial results, and a rich pipeline to prevail over cancer via immunotherapies. It includes an extensive gobbledygook of acronyms like ADCs, BiTES, CAR-T, CAR-NK, TCEs, and TRiTES that needs to be broken down. The purpose of this edition of Ground Truths is to take stock of the progress that is being made and where we are headed. In short, we’ve come a long way!

The takeaway lesson is we can’t have enough in the armamentarium of ways to amp up the immune response to cancer. If we’re going to achieve long term success and cures, in many patients it will take more than 1 drug class within the immunotherapy umbrella, i.e. combinations to get us to the goal. Especially promising is the future use of the immune system to prevent cancer in people at high-risk.

Here’s an infographic summary I made with help from ChatGPT Images 2.0 which, from bottom up, guides the inventory.

And a Table to summarize the number of agents in each category that are FDA-approved or in the pipeline.

In this post, in order to stay on point, I won’t get into some important details for each type of immunotherapy, such as the serious side effects that can occur, like cytokine release syndrome and other sequelae from unleashing the immune system including immune-related colitis, myocarditis, lung toxicity, and hepatitis. Rather, the goal of this summary is to give you a sense of how the field is evolving and its future directions.

Before getting into each category, just to point out there are over 2,500 drugs and programs for cancer immunotherapy now!

At the Foundation: Impact of Immune Checkpoint Inhibitors (ICIs)

The extension of survival (vs placebo) for these drugs ranged from a couple of months to about a year and a half, with extensive variability between types of cancer, different baseline treatments, and between individuals. Overall, the most extensive data available for the PD-1 and PD-1 inhibitors is with Keytruda (pembrolizumab), which now has >40 FDA-approved indications and is the top selling oncology drug in the world. Yet incompleteness of impact of ICIs, with only short-term survival improvement, reflects resistance manifest or developed during therapy. As recently reported, short term fasting (16-hours) enhanced ICI effects in patients. But that doesn’t change the big picture of only partial effects of ICI for most patients. That’s why there’s a whole new crop of ICIs well into clinical testing (TIGIT, TIM-3, VISTA, B7-H3, BTLA, and more) which may help in combination to delay or prevent resistance. For example, the LAG-3 targeted drug was approved in 2022 in combination with a PD-1 inhibitor (nivolumab) for metastatic melanoma. But the ICIs, including the new ones and the combinations, were just the beginning of a superabundance of ways to leverage the immune response against cancer.

I put oncolytic virus treatment as a way to rev up the immune system just above that (Figure above), with less aggressive pursuit than ICIs or the others we we’ll subsequently review. The overall track record looks good for safety, with the virus penetrating and killing cancer cells, along with pushing the immune system to kick in. Its first US approval (T-VEC) was in 2015 vs unresectable melanoma, and there are many viral agents (←an excellent review) in clinical development but they require injection into the tumor.

Share Ground Truths

Antibody Drug Conjugates (ADCs)

As seen above, the antibody bound to drug gets in through the tumor microenvironment to bind to tumor proteins (antigens) on the cancer cell surface. Then the cancer cell killing payload is released.

The are over 15 ADCs approved with many new indications being pursued, as indicated in the Figure below and this outstanding new Cell review. There are now over 300 investigational ADCs in the pipeline. And combination of ADCs with ICIs, presumably additive or even synergistic because of their different mechanism of action, also has initial clinical trial experience.

Bispecific T-cell engagers (BiTES)

BiTEs use the antibodies to bridge a tumor cell surface antigen and a T cell (CD3 on the surface) that rapidly activates cancer cell killing. Schematic below from Dana Farber.

There are already 9 different BiTES approved by FDA for several types of blood (liquid) cancers (e.g. multiple myeloma, lymphoma) and uveal (eye) melanoma. There are well over 500 more BiTES in the pipeline! These include many different T cell surface proteins besides C3 and moving onto tri-specific antibodies (TRiTES). While off-the-shelf is an advantage, they rely on engaging functional, non-exhausted T cells. Like ADCs, they are being explored as combination therapy with ICIs and as pre-therapy for CAR-T to achieve tumor debulking. The 3 generations of the T cell engagers is shown below and reviewed here.

Share

The CAR and TILs Story

I have previously posted on engineering T cells at Ground Truths for cancer, and for its mirror image—autoimmune diseases. A review in April 2026 provides an excellent state-of-the -science for “adoptive” CAR, whereby T cells are isolated from the patient, expanded, engineered with a specific receptor target, and given back to the patient (autologous). The first approval for this cell therapy was in 2017 for blood cancers (e.g. pediatric leukemia) and the success has extended to multiple myeloma and lymphomas but solid cancers have proved to be a formidable challenge. That is attributed to difficulty for getting cells into the tumor and heterogeneity of the antigens within the tumor. Back in 2024. tumor infiltrating lymphocytes (TILs), which use unmodified tumor cells that target many antigens, were FDA approved for melanoma. There are extensive strategies evolving to achieve success in solid tumors, including use of natural killer (NK) cells instead of T cells, and more recently CAR-macrophage (←good review here), and different antigen targets (such as CD70) of engineered cells (Figure), innovative ways to avoid T cell exhaustion, setting up charging stations in the body, or secreting proteins to enhance antitumor action, some referring that concept to an “armored CAR. “ AI is being used to determine best ways to optimize CAR cell binding. Recently (reported April 30), the first use of a stem-cell memory (CAR T SCM) type of of long-lived immune cells given to 11 patients with refractory blood cancers, which achieved success in many. Adding to this work was the use of a cytokine fusion scaffold (interleukin 7, 15 and 21) to augment functionality of the CAR T SCM cells.

While all these tweaks are being pursued, the major anticipation is to move from ex vivo to in vivo—inside the body with allogenic, off-the-shelf CAR therapy. This would take out huge expense and weeks of delay, preempting the need for manufacturing individualized cells manufactured and returned to the patient. Instead better delivery systems (viral and mRNA-lipid nanoparticles) are required, and initial validation has been seen for this in patients with refractory multiple myeloma. The strategies are getting increasingly creative and sci-fi like. A recent cover article (Figure) depicts the use of red blood cells carrying mRNA-lipid nanoparticles (little blue spheres on the red blood cells) that home to the spleen and are phagocytosed (eaten, taken up) by white blood cells (large blue cells) to deliver CAR. None of the in vivo, off-the-shelf CAR have been FDA approved to date, but this strategy is expected to be the mainstay of cell therapy for cancer in the future.

Share Ground Truths

The Vaccines for Treatment and Prevention

There are 2 types of cancer vaccines categorized by their goal: treatment and prevention. Mounting experience has been seen for treating refractory cancers, as summarized in the Table below, with impressive results for many refractory cancers that include pancreatic, glioblastoma, renal and melanoma. Recently, individualized mRNA vaccines were also shown to induce durable T cell immunity for triple negative breast cancer.

Most of these are personalized neoantigens, that is vaccines developed from the patient’s tumor antigens, which takes about 2 months to produce, and often given in combination with an ICI, as seen below. Many use mRNA-nanoparticles for their delivery. Increasing attention to simplify this process is being given to using antigens that are not personalized, to allow for scaling of this approach with reset to timing and cost, along with other ways to rev up the immune system in the vaccine such as adjuvant enhancers. There’s also the concept advanced for putting such vaccines directly in the tumor.

The other category is known as interception vaccines, intended to be used in patients at high-risk hereditary cancers, such as with Lynch syndrome or BRCA mutations. A Phase 2 clinical trial for Lynch syndrome carriers with an off-the-shelf neoantigen vaccine (antigen derivation shown in schematic below) was published earlier this year, demonstrating a robust immune response in 45 participants, given in combination with an ICI. Such work is extending to individuals with BRCA 1/2 pathogenic mutations which could ultimately prevent the need for bilateral mastectomy and risk of other cancers. Beyond that, preclinical evidence for interception vaccines has just been demonstrated vs pancreatic cancer using KRAS as a target.

Where Do We Go From Here?

With about 8 different classes of cancer immunotherapies that are available or under investigation, this is evolving like the multiplicity of treatment choices we currently have for hypertension or Type 2 diabetes that have a double digit number of treatment classes. How are we going to choose the right therapy or combination to achieve best outcomes, reduce side effects and cost? When should immunotherapy be given as the first line of treatment instead of traditional chemotherapy? How reliable are biomarkers that are typically used— PD-L1, microsatellite instability high (MSI-H), mismatch repair deficiency (MMRd), and high tumor mutation burden (TMB)— in guiding therapy? Or mutations in STK11 and KEAP1 genes that may denote resistance to immune checkpoint inhibitors? Do we need to rely on toxic chemotherapy (or surgery or radiation) to de-bulk a cancer before these treatments are applied?

We are missing two fundamental assessments to be much smarter for directing these therapies. For one, we have no immunome in the clinic—a way to assess the functional status of a person’s immune system. I recently reviewed in Ground Truths the landmark papers on AI of the thymus gland. We can derive a thymus health score from a low resolution chest CT scan and this score informs response to immune checkpoint inhibitor therapy, as shown below. It was far better than the PD-L1 biomarker, which is traditionally used for predicting ICI response. Look how survival was influenced for low thymic health scores in the right hand panel below. For decades we have ignored the importance of the thymus in adults for promoting our immune system health, and now have a way to assess it.

There are other ways to assess a person’s immune system, such as a proteomic clock to see if it is outpacing chronological age, and specific proteomic immune cell clocks. These are not yet clinically available but they would fulfill a major unmet need and could help anticipate failure of monotherapy (e.g. using ICI alone or at all).

The other missing piece is our inability to non-invasively assess the tumor microenvironment (TME). The TME consists of tumor cells, matrix, stromal (supportive, framework including blood vessel) tissue and immune cells. If we knew what was the status of the immune cells in the TME of a patient’s cancer we’d have far more insight. For example, high cytotoxic killer cells (CD8+) in TME is a good sign, whereas high proportion of Treg cells is a poor prognosis sign. The goal of therapy is to change a “cold” tumor into a “hot” one (a simplistic Figure below from Dana-Farber) with the latter reflecting high infiltration of T cells, especially those rich with recognition of specific tumor antigens. Currently the use of TME in the clinic is fragmented; there is no standardized, universal TME score or reporting from path tissue. No less non-invasive.

That deficiency just changed. Another landmark paper in the field was published by Aaron Newman and colleagues in Nature providing a non-invasive assessment of TME via a blood sample (“liquid biopsy” and classifying that with machine learning into “spatial ecotypes” (SE) which predicted response to ICI immunotherapy. Note how both of these landmark papers (thymus and TME) relied on AI to crack the case. Below is a Figure from the TME paper for patients with melanoma and response to ICI therapy according to the 9 different SE profiles. I’ll be talking with Prof Newman soon (June 3rd) for a Ground Truths podcast. Please join us.

The importance of the TME can’t be emphasized enough. A paper this week looked into the issue of pancreatic pre-cancerous lesions. We have over 100 of these PanINs in our pancreas (some people have >1,000) and it wasn’t known why only a very few percent of people have these progress into pancreatic ductal adenocarcinoma (PDAC, Figure). The answer: TME.

I don’t know about you, but the thought of having so many pre-cancerous lesions in our body doesn’t sit well. Especially if our immune system is not protecting us like it should. That’s why the new pre-clinical report of a vaccine that is KRAS-directed, preventing pancreatic cancer, is especially intriguing.

That brings me to the future of preventive or interceptive vaccines. They are now being investigationally applied for people with very high-risk mutations (disease-causing, commonly referred to as pathogenic). But the case for using interceptive vaccines will undoubtedly broaden over time. A recent paper on tumor evolution brought out the concept of directly addressing tumor promoter clones, where the risk in the patient has become much greater, calling for “promotion-preventive” strategies (using the term “promolytics”). Enter interceptive vaccines as a prospect.

Now consider that a person has a poor thymic health score or accelerated aging of their immune system clock. Or by genomic assessment they have cancer predisposition gene variants (not pathogenic mutations) and/or a high polygenic risk score for cancer. Or worrisome CHIP clones that we don’t assess.. This sets up identifying a high-risk group as candidates for interceptive vaccines, whether they are generic to amp up a person’s immune system, or specific to the tumor promoter clones identified by blood testing (liquid biopsy). It’s a window for how we can use all the remarkable process in cancer immunotherapy for wide scale immuno-prevention going forward. I hope I’ve convinced you the cancer immunotherapy field is rich and “hot,” and there’s so much more in store now that we recognize our immune system is central to fighting cancer and we are starting to have ways to assess a person’s immune system before they get cancer. Or to help decide which therapy has the best shot.

NB This post was written by me, no AI. As noted, one image was made with the help of AI. I have no COI related to the content of the post.

A big thanks to Ground Truths subscribers from every US state and 212 countries. Your subscription to these free essays and podcasts makes my work in putting them together worthwhile. If you’re not a subscriber, please join!

If you found this interesting PLEASE share it!

Share Ground Truths

Paid subscriptions are voluntary and all proceeds from them go to support Scripps Research. They do allow for posting comments and questions, which I do my best to respond to. Please don’t hesitate to post comments and give me feedback. Let me know topics that you would like to see covered.

Leave a comment

Many thanks to those who have contributed—they have greatly helped fund our summer internship programs for the past two years. It enabled us to accept and support a record number of 51 summer interns coming in 2026! These are high school, college and medical students selected from thousands of applicants. We couldn’t do this expanded program without the funds coming in throughGround Truths.

If you are interested, 2 recent interviews posted in the past week

—The age of prevention

—Extending healthspan

Over the years at WSJ, I relied on Joanna’s reviews of technology for many purchases (and things I avoided) reflecting her keen and brutal assessments. I also had fun working with her on some of her video assessments of health technologies.

Her book I AM NOT A ROBOT is both hilarious and highly informative. It is a terrific primer for those who are not fully grounded in AI getting into the history of AI and right up to date with generative AI’s progress.

To get a sense for why AI thought she was the next Tina Fey, check out the back cover! No surprise she got me laughing hard throughout the conversation, no less while reading the book.

Subscribe now

Here is a short YouTube video she posted on her All In AI year of 2025, referred to during the podcast.

Much of her testing of >100 AI products related to medicine and health care, which is what we especially got into during our conversation. She has 3 major rules in her assessment of AI: (1) Ruthless testing; (2) Benchmark vs human; and (3) Costs, which include “compute cost” that we discussed.

Here are some of the things we covered:

1. AI of her blood work

2. AI of mammogram and breast ultrasound, and overall experience as a person with increased risk of breast cancer

3. AI at the dentist (and ”Dentist Deep Clean”)

4. Assessment of Dr. GPT

5. Bill Gates on health AI

6. Her AI Trainer (Chris)

7. Waymo vs Uber

8. Her AI Therapist (Ash)

9. Nothing Bot Sex (her 2 digital lovers)

10. Impact on her children (including 6 Fucking Hamsters, the robodog Sirius)

A big thanks to Ground Truths subscribers from every US state and 212 countries. Your subscription to these free essays and podcasts makes my work in putting them together worthwhile. If you’re not a subscriber, please join!

If you found this interesting PLEASE share it!

Share Ground Truths

Paid subscriptions are voluntary and all proceeds from them go to support Scripps Research. They do allow for posting comments and questions, which I do my best to respond to. Please don’t hesitate to post comments and give me feedback. Let me know topics that you would like to see covered.

Leave a comment

Many thanks to those who have contributed—they have greatly helped fund our summer internship programs for the past two years. It enabled us to accept and support a record number of 51 summer interns coming in 2026! These are high school, college and medical students selected from thousands of applicants. We couldn’t do this expanded program without the funds coming in through Ground Truths.

Thank you , , , , , and >400 others for tuning into my live video with ! Join me for my next live video in the app.

Get more from Eric Topol in the Substack app

Available for iOS and Android

Get the app

Subscribe now

The DL Era of Medical AI

Beginning in 2018, we started to take note that the retina fundus photo, or optical coherence tomography (OCT), had far more information embedded in it when “seen” by AI that was not detected by ophthalmologists. This led to the realization that supervised learning of hundreds of thousands to millions of images yielded superhuman vision. Many papers were subsequently published about AI of the retina image that supported it as a gateway to nearly every organ of the body, no less to the risk of Parkinson’s and Alzheimer’s disease many years before any symptoms appeared. These were each derived from single deep learning models probing the ability for the image to predict a certain condition. I gave a TED talk about this in 2023. A few years ago, Pearse Keane and colleagues published the first retina foundation model, a generalized dataset with many downstream tasks, called (RETFound) from 1.6 million images, demonstrating its prediction of heart disease, stroke, glaucoma, and Parkinson’s disease. This week, a new retinal image foundation model (Reti-Pioneer) from over 100,000 photos was published, adding thyroid disease, gout, and osteoporosis to the long list of conditions, beyond previously established Type 2 diabetes, hypertension, and hyperlipidemia, for determination of level of risk (summarized in the Figure). Who would have thought a retina image was this rich?!

Most people get an eye exam each year or every other year, which typically includes a photo of the retina. More than half Americans had an eye exam last year, or well over 100 million people. Yet none of the extraordinary progress that has been published for superhuman vision of the retina has been incorporated into routine medical practice!

Despite the lack of general use of AI of the retina, at least 4 companies have moved ahead to provide specific results for patients, although sparsely accessible or used in the United States. There’s Optain for cardiovascular risk, Toku Eyes for chronic kidney disease risk and biological age, Mediwhale for cardiovascular and kidney risk, and i-Cognitio Sciences (based in Hong Kong) for Alzheimer’s risk. Although the accuracy of risk of the various conditions is not 100% (= area under the receiver operating characteristic, AUROC of 1.0) this information could be provided for >15 conditions at a nominal cost ($1) or free to all people having an eye exam. But lack of any orchestration of implementing this huge body of work and evidence and issues such as reimbursement have held it back. It’s software that could and should be easily be applied to every retinal image, and someday ought to be part of a “medical selfie” that we could do with automated capture of the fundus using our smartphone and an app readout.

Last week, an AI to detect pancreatic cancer was published which detected ductal adenocarcinoma (PDA) up to 3 years ahead of radiologists (median advance interval 475 days) and the detection of occult PDA almost doubled for AI compared with radiologists (73 vs 39%, respectively, Figure below). The multicenter study also had external validation. Yesterday, the FDA gave approval for access to an experimental drug for pancreatic cancer with very promising results. How about access to an AI to detect pancreatic cancer before it would be picked up clinically?

In China, AI detection of pancreatic cancer via chest and abdominal CT is becoming routine using their validated AI tool known as PANDA. That is to say, the AI is automatically used for detection even though the scan was ordered for something else. That is essence for what has become known as Opportunistic AI. Some examples are shown in the Figure below.

There is a long list of missed imaging opportunities that I’ve summarized in the Table. Just a few weeks ago, as I wrote in a Ground Truths, there’s the ability to obtain a thymus health score via AI of a low resolution chest CT. To get critical information about our immune system health which is not clinically obtainable. We’re leaving so much valuable medical information on the table by not incorporating validated means of AI detection to our medical scans. That doesn’t even take into account what is encoded in electrocardiograms, pathology slides, and many other types of medical images that we’re not extracting with AI.

Share Ground Truths

The LLM Era of Medical AI

In the April issue, the editors in Nature Medicine published this essay calling for evidence. It reminded me of the famous Cuba Gooding line “Show me the money” from the Jerry Maguire movie.

As opposed to many randomized trials and prospective assessment with external validation of medical image AI, the evidence with generative AI in medicine is lacking. That has not held back the public or physicians from using AI chatbots and LLMs. For the public, according to multiple surveys, 12% of adults are using an AI chatbot every day, or about 40 million people, and estimates range widely from 32% to 73% for having used a chatbot for health information in the past year. According to a recent survey by the American Medical Association in March 2026 of 1,700 doctors, 72% are using genAI for at least 1 use case, and 35% for direct patient care, that is non-administrative, decision making. The graph below extrapolates that for ~1 million physicians.

The next graph indicates the projected use by physicians (from the AMA survey) by the end of 2026. You can see expectations are far different than current use! Note the plan for diagnosis assist.

For AI based decision support, such as making a diagnosis or optimal management of patients by clinicians, what is the evidence? We don’t really have any real-world data. This week in Science Magazine multiple (6) experiments assessing 2 LLMs (Open AI’s o1 and ChatGPT 4) and physicians for case vignettes supported the potential of improved reasoning and, in a simulation of real world emergency room decisions at 3 touchpoints (graph right below) compared with 2 doctors, there was improved initial triage decisions for 01. Most of the many publications use case studies, simulations, and actors as patients. Hardly representative of the messy world of the practice of medicine.

Share

On the patient side, there’s the same deficiency. A limited exception was a small, prospective, single arm study of 100 adult patients (preprint published comparing a model (AIME) vs primary care physicians for differential diagnosis and management plan, which were shown to be on par. A study of ChatGPT Health for helping to triage patients (simulated) to stay home or go to the emergency room did not fare well for the AI, with many blatant triage errors for real emergencies (such as diabetic ketoacidosis or impending respiratory failure (Figure).

In a randomized trial of LLMs comparing patients assisted by LLMs for 10 medical scenarios (not real-world), the patients did’t do well (Figure) and the conclusion was “We recommend systematic human user testing to evaluate interactive capabilities before public deployments in healthcare.”

There are a couple of exceptions to the lack of real world assessment including a Nature Health paper for use of an LLM in 16 primary care clinics in Kenya and a single center randomized trial in eye care. The latter found a higher diagnostic rate for ophthalmologists using AI (92.2%) compared with those who didn’t (75.4%).

In summary, there is very little evidence for LLMs benefiting patients or doctors for health outcomes. That is not to say that generative AI doesn’t help. It offers strong support for administrative work, such as summarizing charts for doctors, or reviewing labs for patients, or providing relevant publications to clinicians and patients, or helping with pre-authorizations for doctors and billing questions for patients. AI chatbots are helping patients prepare for visits, understand their diagnosis and prescriptions.

But going back to the Nature Medicine’s call out for evidence, we need it. For LLMs to help doctors make critical diagnoses or treatment decisions, that requires prospective studies, ideally randomized, but at the very least rigorous and large, with independent adjudication of health outcomes. And for the public, how do we know entering all one’s data leads to the correct diagnosis or treatment? This week we learned that the symptom quality reported to an AI was deficient as compared to that presented to a physician. How does the benefit of patient access to genAI compare with potential harm?

Summing Up

I’ve reviewed a major paradox whereby AI for medical images, with extensive research dating back more than a decade ago, is not being implemented. Whether it’s a mammogram, a CT scan, a retinal image, or colonoscopy, that have all been extensively studied, their value to improve accuracy and risk assessment in medicine is being missed and essentially disregarded.

On the other hand, tens of millions of Americans are using AI chatbots for medical support, as are a substantial proportion of physicians. There are many reasons to use AI here that are easy to support, because they represent an extension of a web/Google search. Just with much more specificity and depth of response, not something that would be subject to regulatory oversight. But when it comes to making a diagnosis or providing a treatment plan there needs to be proof that LLMs are improving accuracy and outcomes. We’ve already seen multiple studies (again not real world) when the AI performed better for various tasks than the doctor with AI, including the new Science paper this week, indicating we don’t even know yet the optimal way of deploying AI (the human-in-the-loop question). As Raj Manrai wrote in his excellent explainer thread, as one of the senior co-authors of the Science paper: “What do our results actually call for? Prospective clinical trials. Health systems investing in infrastructure now. Monitoring frameworks that track not just diagnostic accuracy but safety, efficiency, and cost. The science has reached a point where trials are justified.” We can’t get to high performance medicine, relying on generative AI for key decisions, without that.

One sticking point. Unfortunately, by the time peer review papers are published, the models assessed are outdated (such as 01, which is text-only, when GPT5.5, which is multimodal, would be current). That can give the AI enthusiasts cover, saying the lack of optimal AI performance was because of a weak and old model. The reality, however, is to prove it. Publish it quickly as a preprint.

We’re just a couple of years into the LLM era for medicine. Waymo started in 2009 and it took more than 15 years of rigorous, iterative work to show its true superhuman performance for outcomes with >90% reduction of serious accidents compared with human drivers. Let’s fix this paradox of medical AI implementation. It’s a two-fold and major undertaking. Amping up the use of medical AI where it’s proven and performing the clinical trials required to justify wide-scale adoption where pivotal evidence is lacking.

NB This post was written by me, no AI. Two images were made with the help of AI, as indicated. I have no COI related to the content of the post.

A big thanks to Ground Truths subscribers (> 205,000) from every US state and 212 countries. Your subscription to these free essays and podcasts makes my work in putting them together worthwhile. If you’re not a subscriber, please join!

If you found this interesting PLEASE share it!

Share Ground Truths

Paid subscriptions are voluntary and all proceeds from them go to support Scripps Research. They do allow for posting comments and questions, which I do my best to respond to. Please don’t hesitate to post comments and give me feedback. Let me know topics that you would like to see covered.

Leave a comment

Many thanks to those who have contributed—they have greatly helped fund our summer internship programs for the past two years. It enabled us to accept and support a record number of 51 summer interns coming in 2026! These are high school, college and medical students selected from thousands of applicants. We couldn’t do this expanded program without the funds coming in throughGround Truths.

For those of you interested in extending healthspan, my NPR segment this week

“You are a slightly different genetic version of yourself today from yesterday, and will be different yet again tomorrow.”—Roxanne Khamsi

Roxanne Khamsi is one of the leading life science journalists, a contributing writer at The Atlantic, recognized with multiple awards for notable publications. Her new book is entitled BEYOND INHERITANCE. It tells the story about us all being genetic mosaics, chock full of somatic (acquired) mutations, and the implications of those mutations for our health. The myth of a single genome, carbon copy, master blueprint, but instead a dynamic, shifting mosaic in constant internal evolution.

Share Ground Truths

And here’s the back cover. I was delighted to endorse this book, and reread it to prepare for our podcast. It’s extraordinary and mind-bending.

Here are the topics we covered:

—The Math. 330 billion cells of our ~37 trillion turnover each day, which yields trillions of mutations per day.

—Cellular competition. Winners and losers of an “endoevolution,” Darwinian selection inside us whereby healthy or super-fit mutated cells can crowd out the unfit ones.

This was theorized in 1881 in a book THE STRUGGLE OF PARTS, by Wilhelm Roux

which led to Friedrich Nietzsche’s famous quote “Uniformity is pure delirium.”

Subscribe now

—Single-cell sequencing. How this field catapulted forward owing to the ability to zoom in on genomic mutations at the cellular level.

—Cancer chemotherapy overkill. The routine scorched earth, carpet bombing approach can promote resistance and deleterious mutations, leading to an adaptive strategy of leaving some cancer cells behind, as has been shown to be effective in prostate cancer for improving survival.

—The immune system somatic hypermutation. B cells have the theoretical capacity to produce 1 quintillion unique antibodies (a million trillion). If this weren’t possible, we could die from a common cold. In the Covid pandemic, even before the Omicron variant appeared, Covid booster shots induced hundreds of unique antibodies with neutralization capacity against Omicron. The anticipatory “Red Queen” effect.

—Autocorrection of mutations. They can revert, cure a rare genetic disease from within. It can be considered “Natural Gene Therapy.” Example below of a skin condition Epidermolysis Bullosa with revertant mutations (→ normal skin appearance)

—Different role of acquired mutations through the lifespan. Examples: At embryonic stage, Lines of Blaschko (Image below). In mid-life endometriosis, and in older adults Loss of the Y chromosome (LOY) and Clonal Hematopoiesis of Indeterminate Potential (CHIP). Note these 2 somatic mutation clone conditions are associated with risk of diseases; CHIP-cardiovascular and cancer; LOY-heart failure and Alzheimer’s. I’ve written about CHIP extensively here and here.

—Phenocopy. How a somatic mutation can look the same as a germ-line, inherited mutation, with respect to a disease, and how that is determined.

—Environmental effects inducing somatic mutations: UV light, air pollution, plastics

Share

—and 3 new papers in the past week!

1. Somatic mutations in the microglia cells of the brain, same as cancer mutations, drive inflammation and are enriched in the Alzheimer’s brain

   

   2.. The potential of “promolytic drugs” to be used to prevent cancer in people who exhibit precancerous somatic mutations

   

   3. How somatic mutations can be the basis of autoimmune diseases

We also spoke about the role of somatic mutations in aging and super aging

A related excerpt at The Atlantic

A Quick Poll

Thank you , , , , , and more than 500 others for tuning into my live video with ! Join me for my next live video in the app.

And a big thanks to Ground Truths subscribers (> 205,000) from every US state and 212 countries. Your subscription to these free essays and podcasts makes my work in putting them together worthwhile. Please join!

If you found this interesting PLEASE share it!

Share Ground Truths

Paid subscriptions are voluntary and all proceeds from them go to support Scripps Research. They do allow for posting comments and questions, which I do my best to respond to. Please don’t hesitate to post comments and give me feedback. Let me know topics that you would like to see covered.

Leave a comment

Many thanks to those who have contributed—they have greatly helped fund our summer internship programs for the past two years. We just accepted 51 interns for the 2026 summer, a new record, from thousands of high school, college, and med school student applicants. Our fully loaded cost to do this program is ~$300,000 per year. We’re deeply appreciative for your support of Ground Truths that has enabled this program to prosper and expand.

One more thing:

SUPER AGERS was featured on the CBS Morning Show this week in a segment with Dr. Jonathan LaPook, Chief Medical Correspondent, on the meaning and measurement of biological aging.

Get more from Eric Topol in the Substack app

Available for iOS and Android

Get the app

This is one of my favorite books over recent years. Sebastian Mallaby is the Paul A. Cocker Senior Fellow for International Economics at the Council of Foreign Relations and author of 6 bestselling books.

THE INFINITY MACHINE tells the story of AI’s progress over the past 15 years largely, but not exclusively, from Demis Hassabis as the protagonist and leader of DeepMind’, with its 2010 mission statement to achieve superintelligence by 2030. It’s a rich, informative, page turner.

Subscribe now

What We Discussed:

—What is an Infinity Machine?

—Influence of Claude Shannon’s Information Theory and Douglas Hofstadter’s Pulitzer Prize winning book Gödel, Escher, Bach

—Origin of DeepMind in 2010. Prescient. Charter, business plan, included use of agents. How Demis Hassabis was made for the mission!

—Contrasts with Sam Altman and the other AI leaders, the Oligopoly (cover of The Economist this week). For example, Nature papers vs white papers on company websites.

—In March 2016, the same day when DeepMind’s AlphaGo beat Lee Sedol, Hassabis says it’s time to do protein folding (later known as AlphaFold).

—Symbolic AI (historic, deductive, rule-based) vs Deep Learning (Toronto tribe) and Reinforcement Learning (Alberta tribe).

—The Big Miss: DeepMind’s lack of early recognition of the importance of transformer models (leading to ChatGPT), creating a big opening for OpenAI. And why was this missed? The Comeback Story. Is this happening again with coding (not in the book)?

—The AI Arms Race and Hyperscaling

—How the complex relationship between Google and DeepMind evolved

—The Double Cross

—With the dangers anticipated (parallels to Oppenheimer, Manhattan Project, and the atomic bomb), how to promote AI safety?

—Is the major build up of data centers justified?

Thank you , , , , , and more than 600 others for tuning into our live video with ! Join me for my next live video in the app.

And a big thanks to Ground Truths subscribers (> 200,000) from every US state and 212 countries. Your subscription to these free essays and podcasts makes my work in putting them together worthwhile. Please join!

If you found this interesting PLEASE share it!

Share Ground Truths

Paid subscriptions are voluntary and all proceeds from them go to support Scripps Research. They do allow for posting comments and questions, which I do my best to respond to. Please don’t hesitate to post comments and give me feedback. Let me know topics that you would like to see covered.

Leave a comment

Many thanks to those who have contributed—they have greatly helped fund our summer internship programs for the past two years. It enabled us to accept and support 47 summer interns in 2025! We aim to accept even more of the several thousand who will apply for summer 2026.

And, if you’re interested on extending healthspan, two recent podcasts/TV segments on SUPER AGERS

On With Kara Swisher

On PBS

Get more from Eric Topol in the Substack app

Available for iOS and Android

Get the app

Subscribe now

Here’s a quote from Trisha early in the book:

“I’m not a wellness guru. I’m a scientist.”

Another quote from a Vanderbilt professor:

“An effortless, socially appropriately timed bowel movement is the second greatest bodily pleasure known to man.”

Share

The proportion of people with bowel issues is remarkable, a brief summary:

Some of the topics covered in our lively, fun and highly informative conversation:

—The 3 P’s Propulsion, Pliability, Pelvic Floor

—Ten Myths

—99% of farts have no odor; the 1% that do are from hydrogen sulfide, Pepto-Bismol substantially reduces the latter (and an incredible experiment of people with rectal tubes and judges smelling their gas)

—What is the normal frequency of pooping?

—the Bristol Scale of poop appearance and what it means

—Transit time

—The uselessness and danger of bowel cleanses

—Fibermaxxing

—5 minutes max on the toilet and positions to help propulsion

—Psyllium and laxatives

—The art of wiping and bidets

—Fecal incontinence

—Not to miss urges and the TRPV1 and PIEZO2 receptors

—Impact of coffee, alcohol, NSAIDs (Ibuprofen, aspirin, naproxen, etc), and hydration

—What are the different definitions of constipation?

—What gives poop its color?

********************************************

Subscribe now

Thank you , , , , , and >600 others for tuning into our live video with ! Join me for my next live video and newsletters in the app.

And a big thanks to Ground Truths subscribers (> 200,000) from every US state and 212 countries. Your subscription to these free essays and podcasts makes my work in putting them together worthwhile. Please join!

If you found this interesting PLEASE share it!

Share Ground Truths

Paid subscriptions are voluntary and all proceeds from them go to support Scripps Research. They do allow for posting comments and questions, which I do my best to respond to. Please don’t hesitate to post comments and give me feedback. Let me know topics that you would like to see covered.

Leave a comment

Many thanks to those who have contributed—they have greatly helped fund our summer internship programs for the past two years. It enabled us to accept and support 47 summer interns in 2025! We aim to accept even more of the several thousand who will apply for summer 2026.

Get more from Eric Topol in the Substack app

Available for iOS and Android

Get the app

Adrian is a molecular biologist and co-founder and CEO of Genyro, a synthetic biology company. He has published 2 previous books on life science (Life Without Genes and An Intelligent Person’s Guide to Genetics). In this conversation we discussed his new book, cover below and the title of this post.

Subscribe now

ABI. As Adrian puts it, a monumental shift, a second genesis, the “artivolution,” ability to use a molecular Gutenberg press, it’s post-Darwinian.

I made an infographic about the book and our conversation with NotebookLM

The expansion of genome sequences beyond “Fred’s library” (referring to Fred Sanger, father of genome sequencing, little green box)

We talked about the jump from DNA reading to editing and now writing genomes, with the potential of the latter for both good and harm. I disagreed with Adrian about the bright prospects for curing diseases, but there are many possibilities for positive impact, such as intervening vs the climate crisis and sustainability.

We also discussed the large language of life models (LLLMs) and by coincidence my colleagues and I just published a review of these in this Nature Biotechnology (←free access is hyperlinked).

“It should be possible to meet most of humanity’s needs through biologically inspired designs—to solve many global problems, revolutionize health care, extend human lifespan, and create other organisms to order.”—Adrian Woolfson

If you are into life science, I think you’ll find Adrian’s book thoughtful and provocative, even if you don’t agree with some of his optimism like a disease-free organism.

The topic connects with digital biology, a theme that is approached in many editions of Ground Truths, such as this one with Patrick Hsu

***************************************************

Thank you , , , , , and more than 600 others for tuning into my live video with ! Join me for my next live video in the app.

And a big thanks to Ground Truths subscribers (> 200,000) from every US state and 212 countries. Your subscription to these free essays and podcasts makes my work in putting them together worthwhile. Please join!

If you found this interesting PLEASE share it!

Share Ground Truths

Paid subscriptions are voluntary and all proceeds from them go to support Scripps Research. They do allow for posting comments and questions, which I do my best to respond to. Please don’t hesitate to post comments and give me feedback. Let me know topics that you would like to see covered.

Leave a comment

Many thanks to those who have contributed—they have greatly helped fund our summer internship programs for the past two years. It enabled us to accept and support 47 summer interns in 2025! We aim to accept even more of the several thousand who will apply for summer 2026.

Get more from Eric Topol in the Substack app

Available for iOS and Android

Get the app

What I consider as 2 landmark papers in Nature this week (here and here) used AI to quantify health of the thymus in 2 large cohorts, and then correlate the metric to a broad array of health outcomes.

In this edition of Ground Truths, I’m going to cover three questions: (1) How was thymus health determined by AI?; (2) How is thymus health linked to key clinical outcomes? and (3) Are there ways we can promote a healthy (aka rejuvenate) thymus gland in our later years?

Subscribe now

1. How was thymus health determined by AI?

This was accomplished by extensive work developing and validating an AI pipeline for 2 stages, the first for localization and segmentation of the thymus bed, and the second to quantify a digital marker, termed the thymic health score. There’s a lot to this algorithmic development, so I won’t go into all the details, but just provide a rudimentary outline of what was done. The 1st stage used supervised learning with 2 radiologists reviewing 2,461 chest CT scans. That work trained a 3D U-Net model to automatically identify the thymic bed’s 3D cropped region using center of mass coordinates, which reached 99.8% accuracy.

The 2nd stage used a foundation model that was pre-trained with SwAV (which stands for Swapping Assignment between multiple Views), a self-supervised model, that generated a high-dimensional thymus representation of 4,096 features. That compares to the radiology reductionist and subjective scoring of 0 to 3 where 0 means the thymus is fully degenerated and fatty, and 3 is considered dense, intact glandular soft tissue. Notably, the self-supervised learning (SSL) performance was superior to supervised (AUC ~0.75 vs 0.55, respectively) reflecting the shallowness of the 0-3 classifier vs SSL’s holistic, self-taught AI. The thymic health score ranged from 0 to 100, with the highest number indexed to fully preserved thymic health.

The algorithmic work extended to explainability with both Shapely value distribution, meaning the model was holistic not relying on any “magic” pixel, and occlusion sensitivity, demonstrating the model’s performance was not affected by ribs, sternum, lung tissue, but rather focused on the thymic bed directly.

Below are a few saliency maps from the occlusion sensitivity that indicate the thymic bed focus, not affected by neighboring structures, using the jet color scale.

Now the AI was ready for processing ~25,000 CT scans, each over 30 MB files, from the 2 large participant cohorts to provide a thymic health score, it did so in less than 14 hours, which is less than 2 seconds per scan!

Share

2. How is thymus health linked to key clinical outcomes?

The two cohorts were the National Lung Screening Trial, NLST (N=25,031) and the Framingham Heart Study, FHS (N=2,581) each with baseline demographics and long term follow up of >12 years for health outcomes.

From the NSLT, you can see below that increased age and body-mass index were correlated to reduced thymic health. And, overall, men had reduced thymic health scores compared with women.

There are a lot of graphs in the paper for outcomes, for both the NLST and FHS cohorts. To simplify the major outcomes, below is a composite Figure with striking reduction of all-cause mortality and for the different causes of death. The results were consistent between the 2 cohorts. For example, the cardiovascular mortality hazard ratio average was 0.57 in NLST and 0.38 in FHS (weighted to be 0.53 for the aggregate). The reduction of mortality also extended to digestive diseases and pulmonary disease (data not shown below).

Lung cancer incidence was reduced by 36% for high vs low thymic health (Figure below, with a panel -incidence and b panel-mortality, adjusted for age, sex, BMI, smoking status). Notably, there was a smoker’s paradox: lower incidence and improved survival for lung cancer in smokers who had a high thymus health score.

Smoking was associated with lower thymic scores whereas alcohol was not. Higher thymic health was linked to higher HDL cholesterol, lower triglycerides, lower fasting blood glucose, and lower systolic and diastolic blood pressure. Inflammation biomarkers, such as C-reactive protein and interleukin-6, were elevated in the people with lower thymic health scores.

In their second paper, the same team looked at the relationship between thymic score and outcomes for 3,476 patients receiving cancer immunotherapy. The non-small cell lung cancer progression-free survival was reduced in patients with low thymic health score.

A parallel relationship for survival was seen for other types of cancer including melanoma, breast, and kidney with a 44% lower risk among high thymic scores (Figure). A similar pattern was seen with different immune checkpoint inhibitors.

The thymus health score outperformed programmed death ligand-1 (PD-L1)and tumor mutation burden (TMB) assay of the tumor and was an independent prognostic marker of progression-free survival. In this dataset, a direct connect with thymus score and adaptive immune function was noted with a correlation before cancer treatment for both T-cell diversity and thymus T-cell production (a metric known as T cell receptor excision circles).

Thymus Removal Study

There is a highly relevant citation about thymus function before moving onto ways to rejuvenate it. In 2023, a very important study of thymus removal during cardiothoracic surgery was published, with 1146 patients undergoing thymus removal vs 1146 matched controls (mean age 55 years). The adverse outcomes for thymus removal were striking, with a 2.9 higher risk of all-cause mortality and doubling of cancer incidence, 1.5 fold increase in cancer mortality, and 1.5-fold increase in autoimmune diseases. The immune function showed marked compromise after thymus removal, as reflected by CD4+ and CD8+ T cells (Figure below)

Share Ground Truths

3. Are there ways we can promote a healthy (aka rejuvenate) thymus gland in our later years?

In recent months we’ve learned a lot about the process of thymus involution and the pathways by which this may be modulated. Thymus involution is primarily due to loss to thymus epithelial cells (TECs) and there are 2 subtypes with different functions: the cortical, responsible for positive selection of T cells, and the medullary, for negative selection. When adipose tissue infiltrates the thymus, thymic adipocytes are pro-inflammatory, knocking out T-cell output. Likewise, systemic inflammation, or “inflammaging,” from smoking, obesity, and chronic stress promote thymic adipocytes. Now, from preclinical studies, we know about key pathways that account for these processes.

Historically, back in 2014, FOXN1 was recognized as a single master transcription factor essential for TEC function, with studies in mice with forced FOXN1 upregulation leading to thymus regeneration. That laid the foundation for potential regrowth of quiescent thymus cells. It took awhile to find pathways to accomplish this goal and to even refute its role as the master regulator.

The Liver-Thymus Axis

Feng Zhang and his team recently showed that injection of an mRNA vaccine to the liver encoding DLL1, IL-7 and FLT3-L in combination, factors known to support thymopoiesis (Figure), successfully enhanced immune function in aged mice. However, the effect was transient.

The FGF21 Story

Two reports in 2025 addressed the thymus gland’s production of FGF21, a growth factor (fibroblast growth factor 21) and its control of thymus involution. This growth factor is also produced by the liver, but it doesn’t have impact on the thymus in aged rodent models. In contrast, The liver’s hepatocyte growth factor (HGF) has been shown to reverse senescent TEC structural changes. Growth hormone stimulates production of insulin growth factor-1 (IGF-1) in the liver and the thymus. Two small studies in human participants (TRIIM, for Thymus Regeneration, Immunorestoration, and Insulin Mitigation) of 6 and 50 people, respectively, treated with combined human growth hormone, metformin and DHEA, suggested the potential of slowing epigenetic aging and improving thymus mass and function. But the combination of drugs, small sample, and lack of controls make conclusions murky.

Increased thymus FGF21 led to increased CD8 T cells in old mice and extended healthspan and improved physical performance. Building on the previous FOXN1 data, ablation of β-klotho, the obligatory co-receptor for FOXN1, accelerated thymic aging. The benefit to FGF21 was confirmed in a companion paper using an FGF21 knock-in mouse model showing thymus enlargement and increased TECs throughout the lifespan. (schematic Figure below). The accompanying editorial for these 2 papers speculated: “The proposed link between age-associated thymus involution and organismal aging is intriguing, and suggests we might seek to achieve systemic rejuvenation by counteracting thymic involution.”

Other Factors

Lessons from the Axolotl (Mexican Salamander)

In an elegant set of experiments of the axolotl (Figure), known for its complete thymus regeneration after total surgical removal, mediators were identified. Surprisingly FOXN1 was dispensable. But midkine, a growth factor, appeared to be the driver, the initiator of thymus cell growth, survival, and repair. Other components that contributed were Postn+ (periostin-expressing mesenchymal niches) and Ccl19, a chemokine protein from the bloodstream.

RANK and RANKL

RANK (receptor activator of nuclear factor κB its ligand (RANKL) have been considered key regulators of medullary TECs. In a study of both aged mice and human thymus cell cultures, their key role was confirmed, with multiple benefits of resorting TEC function, recruitment of progenitor cells, and T cell development.

Subscribe now

Summing Up

This body of work strongly supports the health of the thymus gland as a critical regulator of human healthspan, not just a correlate or link. The new landmark studies are reinforced by the regenerative biology experimental models that have shown, via an array of mediators, that healthspan of aged mice is at least in part dependent on thymus gland function and its critical role for maintaining adaptive immunity. It sure seems that we’d be better off having an immune reservoir of naive T cells instead of a profile in older adults of exhausted, low quality, memory T cells, with a senescent phenotype.

It is striking that we ignored the importance of the thymus gland for many decades. A wake-up call was the results of the thymectomy matched control study reviewed here, with the opposite of high thymic health score improved health outcomes.

The convergence of AI and healthspan here is notable. The exceptional and laborious work for developing and validating (no less explaining) the AI to quantify thymus health set the foundation for probing the connect the major health outcomes in two cohorts with extensive follow-up. We’ve never had a way to meaningfully quantify the thymus gland before, so this represents important leverage of supervised, self-supervised, and transformer models. The study wouldn’t have been possible without AI. And it begs the question as to whether there are ways we can maintain our thymus at a high health score.

But it’s not so simple to rejuvenate our thymus. There are multiple risks including the induction of autoimmunity, increasing the risk of cancer, and inducing a pro-inflammatory state. For example, factors that increase TEC proliferation could compromise the medulllary negative selection process and leak out auto-reactive (self-attacking) T cells. Thymic cells could be induced to proliferate, but we know, for example, that growth hormone and IGF-1 are associated with an increase risk of cancer. If thymus rejuvenation doesn’t clear out the accumulated senescent cells in older adults, then inflammation within the gland could block the production of TECs. Keeping these risks in mind, we now have identified many ways to keep our thymus healthy as we age that undoubtedly will be tested in clinical trials going forward. Ultimately, the benefit-to-risk tradeoffs will be defined.

In the meantime, we have desperately needed an immunome. As I’ve stressed in multiple prior Ground Truths, we have no test in the clinic to assess a patient’s immune status! It was very encouraging for me to get an email from Hugo Aerts, the senior author of the landmark papers in the days after they were published: “I also wanted to say that your Substack on “Why We Need an Immunome” was inspiring for this work. We found it a very compelling perspective and greatly enjoyed reading it.”

Just having an assessment of a person’s immune status via a chest CT scan, with millions of these performed per year, might help in guiding the right immunotherapy for cancer (e.g. more intensive, combinations for those with low thymic health scores) and help define people who are increased risk for age-related diseases of cancer, cardiovascular, and neurodegenerative. From a practical standpoint, with 20 million chest CT scans performed each year in the United States, opportunistic AI-derived thymic health scores could be routinely provided as part of their interpretation.

The main theme of Super Agers is that we now have many layers of data (genetics, proteins, biomarkers, organ clocks that can be integrated via multimodal AI) for us to define a person’s specific risk decades before one of these age-related diseases leads to symptoms, enabling prevention. From the new studies this week, we can add thymus health score as yet another way to understand a person’s risk, a window into their adaptive immune system and healthspan. And perhaps someday we will be able to safely rejuvenate the thymus, or maintain its health through one’s life, and use the thymus health score to monitor.

NB: I wrote this post (No A.I.)

A Quick Poll

*********************************************************************

Thanks to Ground Truths subscribers (now > 200,000) from every US state and 212 countries. Your subscription to these free essays and podcasts makes my work in putting them together worthwhile. Please join!

If you found this interesting PLEASE share it!

Share Ground Truths

Paid subscriptions are voluntary and all proceeds from them go to support Scripps Research. They do allow for posting comments and questions, which I do my best to respond to. Please don’t hesitate to post comments and give me feedback. Let me know topics that you would like to see covered.

Leave a comment

Many thanks to those who have contributed—they have greatly helped fund our summer internship programs for the past two years. It enabled us to accept and support 47 summer interns in 2025! We aim to accept even more of the several thousand who will apply for summer 2026.

Nearly 2 years ago here in Ground Truths (GT) I wrote “The Under-Appreciation of CHIP” calling out the need to take this assay seriously because it conveyed substantial risk information, independent of traditional factors, for cardiovascular disease, cancer, blood clots, and other conditions. If you are interested in this topic (and I hope you will be), I suggest you go back and read that post because I won’t review much of the background here, just a little. In this edition of GT, I’m going to summarize substantial new data on CHIP, review my results, and call for action.

Share Ground Truths

Quick Background

CHIP is the acronym for clonal hematopoiesis of indeterminate potential. It refers to the acquired (somatic) mutation of a blood stem cell that gives it fitness for clonal expansion, with a clone size abnormality cutoff defined as 2% or greater of blood cells, or specifically a variant allele frequency (VAF) ≥ 2% of a white blood cell driver mutation.

CHIP increases with aging: only 2% of people age 50-69, 10% age 70-79, 20% age 80-89, and >39% age 90+. The most common mutations are DNMT3A, about 48%, TET2, about 22%, ASXL1 11%, and a group of others (such as JAK2, TP53, SF3B1) at 19%. The 3 most common are mutations in epigenetic (methylation side chain, chromatin, regulator) genes.

The presence of CHIP is associated with many disease, especially cardiovascular, cancer, and inflammatory conditions, and the risk increases with higher CHIP frequency.

Last month the American Heart Association published a scientific statement about CHIP: “CH is increasingly recognized as a potential driver of cardiovascular disease.”

Below is the key Figure from that review. Besides advanced age there are other risk factors for developing CHIP. Whit CHIP ≥ 2%, the link with higher levels of inflammation and immune dysfunction, in the context of advanced age, referred to as inflammaging and immunosenescence, is seen. As is the association to multiple cardiovascular diseases and blood cancers. While only 0.5 to 1.0% of people with CHIP > 10% will develop a blood cancer each year, the dysfunction of the immune system and a chronic pro-inflammatory state are detected without disease manifestation. In patients with advanced solid cancers, over 25% have CHIP, much higher than age-matched healthy control individuals.

As reviewed in my previous post on CHIP, the cardiovascular disease hazard ratio for CHIP of 1.8 which exceeds smoking, hypertension, and total cholesterol >200 mg/dl (all at 1.4). A new CHIP risk score has recently been proposed that adds to CHIP status by partitioning CHIP level of risk.

Share Ground Truths

New CHIP Findings

I’ll divide these into disease and treatment related reports, and point out this is not an exhaustive review but a sampling of new studies that strongly reinforce the importance of CHIP.

Here’s a Table of some recent disease related findings:

Let me provide a bit of color for these. For improved response to checkpoint therapy for non small cell lung cancer and colorectal cancer, tagged to TET2 mutations, the mechanism was probed below. Note that circulating CHIPs penetrate the tumor microenvironment and here TET2 mutant clones facilitate increased interferon and T-cell response.

But in a different study of TET2 mutant CHIP, in the presence of an aortic aneurysm, the pro-inflammatory macrophage effect was linked to accelerated aneurysm growth.

An important discovery for resilience to CHIP, protecting against its formation and related blood cancers, was the discovery of a gene variant on chromosome 17 in 3 large cohorts (Figure) linked to reduction of multiple subtypes of CHIP (driver mutations) and blood cancers. This turned out to be related to an RNA network and specifically down-regulation of the RNA binding protein MSI-2. This seminal finding could lead to a new intervention to suppress CHIP and its adverse health outcome impact.

Now let’s move on to some new treatment findings.

Low dose colchicine was shown to have a consistent effect of reducing CHIP in a randomized placebo-controlled trial with longitudinal sampling (Table).

This anti-inflammatory drug has previously been studied in the atherosclerotic murine model and patient cohorts with TET2 mutations to reduce plaque and heart attacks, respectively.

Beyond colchicine (an NLRP3 inflammasome blocker), the IL-1β–neutralizing antibody was shown to reduce events in participants with TET2 in a randomized trial and other potential interventions have been reported such as metformin (vs DNMT3A), Vitamin C (vs.TET2), inhibitors to other interleukins such as IL-6, or inflammasome blockers (Figure). The NLRP3 inflammasome inhibitor selnoflast is being assessed in an ongoing clinical trial for CHIP. Even GLP-1 drugs, known to have portent anti-inflammatory action, have recently been raised as a potential effective anti-CHIP intervention. I should mention that low-dose colchicine, which has the best dataset to date, is remarkably cheap, pennies per day, generic, and well tolerated.

Subscribe now

My CHIP Results

In response to my post on CHIP a couple of years ago, a company (Lucence) offered to provide my CHIP results free of charge. I recently had the blood test and got back the results. My CH was 0.3%, well below the 2% threshold for diagnosis and the only driver mutations detected were in TET2 (4 different variants each at very low VAF). Compared with the general population, not adjusted by age, the result indicated a lower risk for the 3 major outcomes as shown below. In my age group, the chance of having CHIP (defined as >2%) is more than 10%, so I feel fortunate. Had I been found to have TET-2 CHIP I might have considered starting low-dose colchicine.

The Implications

We’re learning much more about CHIP, how it links to a risk of many diseases spanning cardiovascular, cancer, thromboembolic (clotting) and liver (previously reviewed in GT). Different driver mutations carry different risks, and even the same CHIP mutations, such as TET2, can do both harm (aneurysm growth) and good (improve response to soldi tumor immunotherapy). The fact that these are multiple clinical studies fully backed up by animal models supports evidence for causality. Some mutations, like DNMT3A, have been linked to improved survival in colon cancer, so the presence of CHIP is not always a bad sign. While JAK2 is an unusual subtype of CHIP (range 0.5 to 4% of CHIP driver mutations), it is specifically associated with clotting complications, such as pulmonary embolism.

The CHIP cause and effect relationship is not yet considered fully established but the new data from multiple parallel human and animal studies strongly support that CHIP is not just a link to adverse outcomes, that it is highly likely to be a driver.

So why don’t we have routine CHIP assays for people of older age? One reason would be we don’t have therapies that are effective, but that seems to be negated with low- dose colchicine and many other candidate drugs with promising data for effectiveness (either blocking CHIP formation or its downstream risk of diseases). Another reason is that CHIP is hard to assay accurately. While it can be detected by whole genome or whole exome sequencing, the gold standard is to use a narrow, targeted panel of ~100 genes and do very deep sequencing, exceeding 500X. That is what was used by Lucence to do my assay, but it has a charge of $950 which is extremely high and I consider untenable. With the marked reduction in sequencing costs in the last couple of years, getting a precise and accurate CHIP assay at low cost should be a priority, in conjunction with dedicated clinical trials of drugs that prove outcome improvements.

Another potential downside of CHIP lab tests is inducing anxiety, such as seen in cancer survivors, and others with high-risk results for major diseases. But the trade-off of close surveillance of such individuals and mitigation of risk factors deserves consideration. Getting ahead of a disease enables prevention, which is typically far better than resorting to a treatment.

With all this background, it is notable that there are currently 2 academic centers that have established dedicated CHIP Clinics, both the University of Chicago and Cleveland Clinic. After my initial post, I learned from Dr. Jennifer Kwan there is a hybrid cardio-oncology CHIP clinic at Yale Medicine. Why aren’t there more? With all the evidence that keeps accumulating, why aren’t we making CHIP assessment widely available at low cost for select individuals (such as 70+ years of age)? Our blood makers for inflammation, such as high-sensitivity C-reactive protein, are notoriously non-specific and can be misleading. I would like to change the CHIP acronym to Clonal Hemaotopoiesis of Important Potential, getting rid of the Indeterminate qualifier. It is certainly informative. I think more so than most of the hundreds of tests that companies like Function Health provide to their several hundred thousand customers. Like Lp(a), will it take another 8 years before we have new guidelines from cardiovascular, hematology and oncology professional societies that back up the utility of diagnosing and treating CHIP?

NB: I wrote this post with no AI assistance. I have no relationship with any company that does the CHIP assay or potential treatments that would treat CHIP.

**********************************************

Thanks to Ground Truths subscribers (now > 200,000) from every US state and 212 countries. Your subscription to these free essays and podcasts makes my work in putting them together worthwhile. Please join!

If you found this interesting PLEASE share it!

Share Ground Truths

Paid subscriptions are voluntary and all proceeds from them go to support Scripps Research. They do allow for posting comments and questions, which I do my best to respond to. Please don’t hesitate to post comments and give me feedback. Let me know topics that you would like to see covered.

Leave a comment

Many thanks to those who have contributed—they have greatly helped fund our summer internship programs for the past two years. It enabled us to accept and support 47 summer interns in 2025! We aim to accept even more of the several thousand who will apply for summer 2026.

Here are some of the topics we discussed and a few related hyperlinked citations:

—Importance of Deep Sleep (and her paper on what happens when deep sleep is purposely disrupted, how it is modulated, effect of alcohol

—The Orexin Antagonist drugs that promote sleep (and a study that shows they can reduce p-tau217 and other neuroinflammation markers

—Link of sleep regularity with less all-cause, cardiovascular, and cancer related mortality

Figure 3.9 from SUPER AGERS (SRI-sleep regularity index)

Share

More things we discussed:

—A sleep foundation model that predicts 130 diseases

—Impact of menopause on sleep

—Getting to sleep vs staying asleep

—Role of naps

—Impact of interruptions of sleep

—Sleep apnea and new interventions

—Vagal nerve stimulation and sleep

—Cerebrospinal fluid wave that occurs during attention lapse after poor sleep

—cognitive behavioral therapy for improved sleep

—Wearables, sleep scores, and effect on sleep

—Any supplements that help sleep?

—The rare genetically endowed short sleepers

Subscribe now

Thank you , , , , , and more than 600 others for tuning into my live video with ! Join me for my next live video in the app.

**************************************************

Thanks to Ground Truths subscribers (> 200,000) from every US state and 210 countries. Your subscription to these free essays and podcasts makes my work in putting them together worthwhile. Please join!

If you found this interesting PLEASE share it!

Share Ground Truths

Paid subscriptions are voluntary and all proceeds from them go to support Scripps Research. They do allow for posting comments and questions, which I do my best to respond to. Please don’t hesitate to post comments and give me feedback. Let me know topics that you would like to see covered.

Leave a comment

Many thanks to those who have contributed—they have greatly helped fund our summer internship programs for the past two years. It enabled us to accept and support 47 summer interns in 2025! We aim to accept even more of the several thousand who will apply for summer 2026.

+++++++++++++++++++++++++++++

I also want to thank the National Academies of Science, Medicine and Engineering for their recent recognition.

Get more from Eric Topol in the Substack app

Available for iOS and Android

Get the app

Subscribe now

How They Are Measured

Cardiorespiratory fitness (CRF) is a real world assessment of a person activities, such as walking or on a treadmill, a reflection of a person’s resting metabolic rate, measured in metabolic equivalent of task (MET) units with 3 recognized levels of intensity : Light (<3.0 METs), example slow walking; Moderate (3.0-5.9 METs), example brisk walking, 3-4 miles per hour; and vigorous intensity (>6.0 METs), example jogging. 1 MET is the energy used in sitting or resting; 10 METs requires 10-fold the energy expenditure. CRF integrates cardiovascular, lung and musculoskeletal functional capacity.

There are multiple methods to calculate your METS, including a standard treadmill MET chart (below left) that plots speed and incline, use a formula if you are doing the Bruce treadmill protocol or the chart below (right), or using heart rate (with any aerobic activity, such as bicycling or jogging) with the formula: METS=0.05 X heart rate+2. So if your HR got to 140 that would be 9 METS. For every increase in heart rate of 10 beats per minute, there’s about a 1 MET increase.

Maximal oxygen uptake (V02 max) is only accurately determined as a performance lab test with a metabolic cart, trained technicians, a specialized tightly fit mask that captures every molecule of inhaled oxygen and exhaled C02 on a ramped treadmill or stationary bike exercise protocol until absolute exhaustion. This is the ceiling of aerobic power achieved via direct gas exchange. A V02 max test costs about $150 for a standard assessment in a university lab.

V02 max by wearables are obviously not measured by gas exchange or directly, but instead through various imputations based upon population-based algorithms of heart rate and movement (GPS/accelerometry). Studies have assessed the Apple Watch, Garmin Fenix 6, and Fitbit with a mean absolute percentage error of 7-16%. Overall, they have been found to consistently underestimate V02 in fit people while overestimate in unfit individuals. They also rely on optical heart rate (which may be inaccurate in people of color), device positioning and wrist anatomy, and can be influenced by such factors as hydration status, altitude, and ambient temperature. Typically, a 6-minute walk is the basis for a wearable to provide the user a new V02 max result. That may not be at all representative of an individual’s exercise capacity.

Share

The Datasets For Assessment

Datasets for Cardiorespiratory Fitness

In JAMA 2009, a meta-analysis of 33 studies of cardiorespiratory fitness was published for the relationship to all-cause mortality in a total of 102,980 participants. A better CRF (per 1 MET higher) was linked with a lower all-cause mortality (Figure) and individuals who had achieved 7.9 METs had substantially less all-cause and cardiovascular mortality. One MET higher CRF was associated with a 14-15% reduction of mortality.

In 2016, the American Heart Association issued a scientific statement on CRF and asserted it should be regarded as a clinical vital sign, reviewing all of the published data to that point.

In 2018, Mandsager and colleagues from the Cleveland Clinic published their data from 122,007 consecutive patients who underwent exercise treadmill testing and had long term follow-up for outcomes.

Here is the Table of METS performance by age and sex. You can see there are 5 categories (columns) from low to elite.

All-cause mortality by sex and the 5 levels of performance are plotted below. The hazard ratio of 1.41 (about 40% increases risk of all-cause mortality) for above average vs below average was the same as the risk of smoking or diabetes. The hazard ratio for mortality from low to elite was more than 5-fold. The favorable impact for women beyond men for METS was seen for each performance group. These results were adjusted for potential confounding variables.

In 2022, Kokkinos and colleagues published CRF exercise treadmill data for over 750,000 veterans aged 30 to 95 years with a mean follow-up of 10.2 years. The analysis was based on 6 categories of MET performance but the hazard ratios were similar to the Cleveland Clinic data (e.g. extremely fit vs lowest 4-fold in this study, 5-fold hazard in the prior one).

Here is a good summary graph of that study. In both there was no risk of increased mortality at the highest fitness strata—in fact it was consistently lower for each age group.

Datasets for V02 Max

There are limited data for direct measurement of V02 max and outcomes. The 2001 Kuopio study from Finland of 1,294 men with 10.7 year follow-up did measure V02 max directly once at baseline along with a symptom-limited exercise tolerance test on a bicycle ergometer. The relationship of V02 max (by quartiles) to all-cause mortality is shown below.

In a 2024 meta-analysis of 42 studies with V02 max (categorized as “objectively measured CRF”) and estimated CRF for prediction or all-cause and cardiovascular mortality, the results were remarkably similar (cardiovascular mortality, 14% reduction, graph below) but notably there were 234-fold more participants with exercise CRF than by V02 max measurements, or >99% of the data is derived from METS. That is to say, nearly all the data we have for link to outcomes comes from CRF, not V02 max.

Reference standards have been published by age group for V02 max. For more information such as by sex, please check the link.

There are other specific studies in heart failure ,chronic obstructive pulmonary disease, pulmonary hypertension and pre-operative evaluation that show use of V02 max can help guide risk or treatment.

Share Ground Truths

Conflation and the V02 Max Craze

The leading proponent of using V02 max in recent years has been Peter Attia, through his podcast The Drive and book Outlive. He has consistently asserted “V02 max is the singular most powerful marker for longevity.” But the problem is conflation. He cites all of the studies of CRF without measuring V02 max and extrapolates to a V02 max result (see side-by-side Kokkinos study Table and Outlive Figure footnote below) and throughout his discussion of exercise in Chapter 11 of Outlive. For example, he writes: “this number [V02max] turns out to be highly correlated with longevity” citing all studies that did not measure V02 max.

In a recent YouTube video by Joseph Everett and Nick Norwitz entitled “Hidden Data: How the Top Longevity Doctor Tricked Us All” there is a segment about V02 max and this significant issue of conflation, discussed by Chris Masterjohn. Below is the relevant 6 minute clip within the longer video. It includes a bit of the 60 Minutes segment with CBS correspondent Nora O’Donnell doing a V02 max text and Peter’s assertion: “We don’t have a single metric of humans that we can measure that better predicts how long they will live than how high their V02 max is.”

As Masterjohn aptly points out, the fixation on V02 max, which is not actually supported by the data, also misses out on our ideal goal of diversity of exercise, including strength and balance training. Indeed, Kim et al, analyzing over 70,000 UK Biobank participants, for both CRF (submaximal bicycle test) and grip strength with all-cause mortality and concluded: “Improving both CRF and muscle strength, as opposed to either of the two alone, may be the most effective behavioral strategy to reduce all-cause and cardiovascular mortality risk.”

Bottom Line

I’ve never done a V02 max and see no reason to do it with the issues of cost, inconvenience, and the pain. As Attia correctly states about going to maximal exhaustion to get V02 max: “If you’ve ever had this test done, you will know just how unpleasant it is.” For spartan, Olympic, high performance athletes who are in high intensity training, or in patients with heart failure or pulmonary hypertension there may be a place for serial V02 max measurements, providing highly objective “goal standard” physiologic metrics. Otherwise, there are no supportive data for people going out and getting a V02 max and making this the focus of their exercise training. That is the reason I didn’t even mention V02 max in my Super Agers book.

Nearly all of the relevant data related to outcomes are based on exercise on a treadmill or bicycle with METS as the index of cardiorespiratory fitness. We should not be placing much value on our smartwatch data. My Apple Watch gave me encouraging high V02 max data over 6 months to suggest my CRF is well above people in my age group (70+, see reference standards above) but I know the data are woefully unreliable.

The problem now, with so much misplaced hype on V02 max, is that most people are using their smartwatch output for gauging their cardiorespiratory fitness, like the 2 patients I mentioned at the top of the post. It’s free to calculate your METS! And that is the real basis of the relationship to all-cause and cardiovascular mortality that has been firmly established in the peer review literature. To be clear, Attia’s assertion is false: "V02 max is the single greatest predictor of your lifespan."

This problem surfaced recently with the introduction of ChatGPT Health. Geoffrey Fowler, the tech journalist at the Washington Post, submitted all his Apple Health and asked for an overall assessment of his health (actual prompt: “give me a single score (A-F) for my cardiovascular health over the last decade including component scores and an overall evaluation of my longevity.”) It gave him an “F.”

Then he gave ChatGPT Health his electric medical record access (portal) and asked the same question again. It gave him a “D’ and attributed that to his V02 max data of 34 ml/kg/min in the past year, below a 45-50 year old male! He also entered his Apple Watch data to Claude Health and it gave him a D+ status Specifically, Claude Health gave him a C- because his V02 max had declined from 41 to 32 ml/kg/min from 2016-2026. But he had over 7,500 steps/day throughout the decade.

These outputs are indicative of the problem—the unreliable wearable V02 max data have become unduly emphasized by current AI platforms using smartwatch data! That will only make the problem worse, adding to the confusion, conflation, and emphasis on the wrong metric.

I hope this post helps to sort out what we know and that the datasets for cardiorespiratory fitness, representing real world physical activity— not V02 max —are the basis of the link to improved survival and freedom from cardiovascular mortality.

If we’re going to focus on a metric it ought to be METS, not V02 max. Not only is it free, simple and universally available, but it is the one best studied for health outcomes. And perhaps the better strategy is to be as physically active as possible and not worry about any metric!

NB: No AI was used in any way to write this post. As mentioned in the caption, I got help from Gemini-3 to produce the first Figure. I have nothing to do, no COI, with any company working with cardiopulmonary fitness or V02 max.

********************************************************************

Thanks to Ground Truths subscribers (> 200,000) from every US state and 210 countries. Your subscription to these free essays and podcasts makes my work in putting them together worthwhile. Please join!

If you found this interesting PLEASE share it!

Share Ground Truths

Paid subscriptions are voluntary and all proceeds from them go to support Scripps Research. They do allow for posting comments and questions, which I do my best to respond to. Please don’t hesitate to post comments and give me feedback. Let me know topics that you would like to see covered.

Leave a comment

Many thanks to those who have contributed—they have greatly helped fund our summer internship programs for the past two years. It enabled us to accept and support 47 summer interns in 2025! We aim to accept even more of the several thousand who will apply for summer 2026.

Subscribe now

1, Improved Accuracy of Detection with Reduced Workload

The National Cancer Institute estimates that 20% of breast cancers are missed by mammograms, leaving plenty of room for the potential of AI to help.

The first study that caught my eye about the promise of AI for mammography interpretation dates back to 2019 from NYU with supervised learning of more than 1 million mammogram images and 14 radiologists. The area under the curve with AI, a performance metric, was 0.895 (1.0 is perfect) and the conclusion was: “We also show that a hybrid model, averaging the probability of malignancy predicted by a radiologist with a prediction of our neural network, is more accurate than either of the two separately.”

Since that time, retrospective studies of various AI algorithms were reported on a frequent basis. Then, in 2023, a real world of medicine study in Hungary with the Kheiron algorithm was published in Nature Medicine. Like MASAI’s design, comparing 2 radiologists vs 1 (or 2 for highest risk of 10 score) plus AI, it found enhanced detection of cancer, 83% of which are invasive.

Notably, there was a >30% reduction of radiologist’s workload. This report from Hungary led journalists to travel there and study their cancer screening clinics, where 5 were performing 35,000 screenings a year with AI since 2021. This article appeared on the front page of the NY Times in March 2023.

Screening in the US is different from European countries because it relies on a single radiologist. That raises the question as to whether the big workloadreduction would also be seen here. But use of AI to partition very low risk scans, a so-called “delegation strategy,” would be expected to achieve a substantial workload reduction in settings with a single radiologist.

With that background, let’s go on to the landmark MASAI trial results, initially reported last year. This trial used the Transpara algorithm, developed by ScreenPoint Medical in The Netherlands, with 2D mammography. Subsequently, on 30 January 2026 the two-year follow up was published. The Table below summarizes the main results for both reports.

From the screening phase, there was a 29% increase in detection of cancer with AI, with a 24% increase in invasive cancer. There was no significant difference in the recall rate. The false positive rate was 1.4% in both arms of the trial. There was a 44% reduction in screen-reading workload.

At 2-year follow up, there was 12% less interval cancer detected in the AI arm, and the cancers picked up by AI were smaller, less aggressive: 16% reduction of invasive cancers and a 27% reduction of non-luminal A subtypes (including triple-negative and HER2-positive cancers). Accordingly, the AI support served a preventative role for earlier detection, catching aggressive breast cancer at the time of screening. But there is more to AI and prevention as we’ll get to in the next section. Of note, the Transpara algorithm is being used in 2 other large randomized breast cancer screening trials: in Norway, a trial of 140,000 women, called AIMS and in the US with 400,000 women, known as PRISM.

Share Ground Truths

Back-up From Real World of Medicine Studies

Besides the prospective, real world study conducted in Hungary, a nationwide study in Germany reported on over 460,000 women undergoing AI-supported mammography reading (N=260,739) or without AI support (controls, N=201,079), with the finding of a 17.6% high rate of cancer detection. It used a different algorithm called Vara MG from a German company.

In the United States, a large study that compared AI-support of a single radiologist (N = 208,891) for what AI identified as “at-risk” (multistage AI-driven workflow) to a radiologist (controls, N = 370,692) there was a 21.6% increase in breast cancer detection rate. The AI used was from DeepHealth. The results were consistent across ancestries and breast density, as seen below.

Another single center, smaller (N-55,581) prospective study from Sweden also showed better detection with AI support, using a different algorithm called Insight MMG from Lunit.

In a prospective, real world multi-center assessment in over 25,000 women in South Korea, the use of AI support led to a significant increase in cancer detection with no rise in the recall rate.

All of the real world studies and MASAI showed no significant increased in the recall rate. That is to say, all of the benefit was accomplished without prompting more callbacks to women for something concerning about the scan.

2. Prevention of Breast Cancer

As you’ve seen above, as validated in the MASAI trial, AI support identified more of the aggressive cancers up front, and at 2-years of follow-up the results reinforced the advantage of increased early detection. But there is another dimension to preventing breast cancer: identifying risk in what is read out as a normal mammogram. in June 2025, CLARITY Breast received the first FDA de novo authorization for predicting risk for developing breast cancer over the next 5 years from a standard 2D mammogram. The original work behind this algorithm dates back to a publication in 2021 in Science Translational Medicine. This approval was based on training from ~420,000 mammograms from 27 facilities in the US, Europe, and South America. Subsequently, verification data from >122, 000 mammograms 10 US health systems, were presented at the RSNA in December 2025. About 16% of women in their 40s with a normal mammogram were found to be high risk. At the recent RSNA, data are presented taking breast density into account: women in the high-risk AI group had more than a 4-fold higher cancer incidence than women in the average-risk group (5.9% vs. 1.3%). But breast density alone showed only small increases in 5-year risk (3.2% for dense vs. 2.7% for non-dense).

Multiple studies have confirmed the ability to use AI for prediction of future breast cancer, such as one report in over 110,000 women and in a smaller study of more than 13,000 women.

In parallel to CLARITY Breast, there has been extensive work by MIRAI, an academic research model initially developed at Harvard-MIT-Mass General Brigham that has been validating risk and prediction for the next 5 years using AI of the screening mammogram. The work has been extensive in over 21 counties and ~2 million mammograms, in diverse populations. At the December 2005 RSNA annual meeting, prediction of interval cancers was assessed for a cohort of over 130,000 women in the UK. The top 20% of risk accounted for 42.4% of the interval cancers.

While MIRAI is not a commercial product, the algorithm for prediction by DeepHealth (owned by RadNet) and Lunit, referred to above, are in the pipeline for FDA review for their prediction capabilities.

AI detection of high-risk sets the patient up for a tighter surveillance plan that could include MRI or ultrasound, multi-cancer early detection blood tests, or simply more frequent mammography (than every other year) with higher resolution tools such as 3D, digital tomographic synthesis, with an established higher rate of detection than 2D mammography. Identification of high-risk could also lead to a workup for genetic predisposition such as BRCA or other pathogenic mutations that were not previously assessed. When positive, that could also initiate a cascade of checking family members for such mutations.

Share

3. Risk of Heart Disease

The leading cause of death in women in heart disease, not cancer. The diagnosis of heart disease is all too often missed in women for various reasons including clinician bias and atypical symptoms.

Breast arterial calcification (BAC) as detected by AI, turns out to be an important marker for risk of cardiovascular events which affords mammography offer a 2-fer, a form of opportunistic AI imaging. In 2023, the FDA gave clearance to CureMetrix cmAngio for AI detection of breast arterial calcification. That was based on a study on over 10,000 women tracking breast arterial calcification over 4-5 years, and found that an increased in category from mild to moderate) was linked to a 59% higher risk of major adverse cardiovascular events , and for moderate increased to severe BAC a 91% increase. A new appearance of BAC from negative was associated with a 41% increase MACE risk.

Other studies have assessed prediction of heart disease in over 49,000 women followed 9 years using a deep learning algorithm with some predictive accuracy.

In another cohort of more than 5,000 women followed for 6.5 years, BAC was linked to a 51% increased risk of cardiovascular disease events. That was the basis for FDA clearance of Mammo+Heart (below Figure is the summary of data from the study)

Some health systems are currently providing BAC data with for a mammogram for free while others charge $90-$120 for Mammo+Heart or up to $75 for CureMetrix. Other companies are working on FDA clearance for BAC, including DeepHealth

The Obstacles and Limitations

Getting an AI read out from a screening mammogram in the United States is restricted to RadNet, a very large network of over 400 locations, the largest radiology outpatient imaging service in the country. But it only operates in 8 states. And their readout adds cost to the patient that is currently $40 out-of-pocket. CLARITY Breast is now getting rolled out in the US but it also has an extra fee attached for its software interpretation of 5-year risk. And you’ve seen how the breast artery calcification also adds more to the cost. These are unacceptable added costs to patients to get the most information from their mammogram. I’ll come back to this below.

There are numerous algorithms for the initial screening but the only one that has been validated in MASAI is Transpara, which is also being used in ongoing large randomized trials in Norway and the US. RadNet uses DeepHealth, Hungary used Kheiron, Germany used Vara, and a center in Sweden used Linut, all of which have not had the same rigorous assessment. It is possible that there are differences in the performance of these algorithms which would not replicate the findings from MASAI, or, alternatively, could surpass its performance. All these algorithms are being refined, too, such as Transpara tested in MASAI, has been substantially tweaked.

The fact that there are ongoing large randomized trials in Norway and in the US to assess mammography either AI support would suggest this matter is unresolved, that we remain in equipoise. One could argue that such trials may no longer be needed, that the overall proof of benefit is clear, that enrolling participants in a control group without AI support might even be considered unethical. But given current implementation issues in the US with very limited access, added out-of-pocket costs, and lack of insurance coverage, the counter to this would be we’re not ready now to change medical practice anyway. Another point put forth is that survival has not been shown to be improved from mammography AI, but it is not hard to connect the dots from the MASAI data on interval cancers. Note there are few randomized trials with over 100,000 participants and major health outcomes.

For CLARITY Breast and other 5-year predictive algorithms, a concern is that there is no standard protocol for what to do when high-risk is identified and lack of proof that the information changes outcomes.

The conclusion from the recent MASAI paper was: “Further analyses of subsequent screening rounds and cost-effectiveness will clarify the long-term balance of benefits and harms and could provide a strong rationale for implementing AI in population-based mammography screening programmes, particularly in the context of workforce shortages.” This is similar to the vast majority of publications that say in the last line that more research is needed, not making a definitive statement.

The models that led to currently approved or cleared mammography AI tools are all derived without generative AI. Given the power of transformer architecture and foundation models, we can expect to see new AI tools that do all 3 tasks above—detect cancer, prevent cancer, and assess risk of heart disease—in one package. That could make all of this even more accurate, faster and cheaper.

When Is It Time to Change Medical Practice?

Let’s go back to my provocative title of this piece, where I assert all mammograms should incorporate AI. In the US, over 40 million mammograms are performed yearly. The whole rationale for mass screening is early detection. Think of the NCI estimate that 20% of US mammograms read out by radiologists miss cancer. In the US we rely on a single reader unlike European countries that use 2 radiologists.

We now have proof that AI support achieves superiority for the early detection goal, but more than that. It is better (~30% enhanced cancer detection) and faster, saving substantial workload on the part of radiologists. Put that in context to the severe and growing shortage of radiologists. It should be cheaper if the software were embedded in the mammography readout without any additional cost to the patient. If used at scale, a small fee of let’s say $1 could be paid to the company providing the AI. Moreover, the cost of treating Stage 3 breast cancer is between $160,000 to $200,000 in the first 2 years so any early diagnosis at Stage 1 or 2, or even better high-risk , could lead to marked savings in cost.

Mammography with AI support has emerged as the most rigorously studied of all possible indications. It is a standout for demonstrating superhuman performance—“digital eyes”— that see things which humans can’t. I believe the data we now have is compelling and should set the stage for a new standard of care. As I’ve reviewed, there are multiple obstacles that are holding us back, but it all goes back to why aren’t we providing all women with the best information available on their risk, early detection, and potential prevention of breast cancer and, no less, that of heart disease? If not now, then when?

NB: No AI was used in any way to write this post. I have nothing to do, no COI, with any company working in mammography AI.

********************************************************************

Thanks to Ground Truths subscribers (approaching 200,000) from every US state and 210 countries. Your subscription to these free essays and podcasts makes my work in putting them together worthwhile. Please join!

If you found this interesting PLEASE share it!

Share Ground Truths

Paid subscriptions are voluntary and all proceeds from them go to support Scripps Research. They do allow for posting comments and questions, which I do my best to respond to. Please don’t hesitate to post comments and give me feedback. Let me know topics that you would like to see covered.

Leave a comment

Many thanks to those who have contributed—they have greatly helped fund our summer internship programs for the past two years. It enabled us to accept and support 47 summer interns in 2025! We aim to accept even more of the several thousand who will apply for summer 2026.