Face Aging Rate uses AI to predict cancer outcomes by tracking biological age changes over time

Face aging rate quantifies change in biological age to predict cancer outcomes - Nature Communications

Aging is a multifaceted biological process characterized by the decline of physiological functions and increased vulnerability to disease and death. It affects nearly all living organisms and is intricately linked with various diseases, particularly cancer, as both result from the accumulation of cellular damage over time. Chronological age has long been recognized as a predictor of survival; this relationship is particularly evident in cancer patients. However, it treats all individuals within an age group identically, disregarding variations in biological aging rates. Recognizing these limitations, it becomes essential to explore biological age indicators that can be readily implemented into clinical practice, especially those that quantify the rate of aging and are easily accessible, for personalized risk assessment. Recent longitudinal evidence demonstrates that significant variation in biological aging trajectories can already be quantified in young adulthood, underscoring the potential for early interventions before disease manifestation. This study aims to determine whether the Face Aging Rate (FAR), a biomarker calculated using artificial intelligence (AI) from facial photographs taken at different radiation therapy courses at distinct time points, is associated with survival outcomes in cancer patients. Facial aging involves alterations in skin texture, loss of volume, and changes in bone structure, resulting in more aged appearances. These visible changes offer a unique window into aging, especially when observed over time. Advances in deep learning have enabled accurate age estimation from facial images by analyzing visual biomarkers, and preliminary studies demonstrate the potential of this biological age estimate from photographs as a powerful prognostic factor. Previous studies have shown that cancer patients often appear older than their chronological age, and individuals with AI-predicted older FaceAge have worse survival outcomes. Combining facial scans with other biometric data has further been shown to be useful for cardiovascular risk assessment. Our previous work introduced Foundation Artificial Intelligence Models for Health Recognition (FAHR-Face), a model trained on over 40 million facial images that capture health-related facial features. The derived FaceAge showed that patients appearing five or more years older than their chronological age had a 21% higher mortality risk. The model revealed significant correlations with health behaviors (e.g. smoking, alcohol, drugs). While single time-point biomarkers provide valuable prognostic information, serial measurements often offer superior insights into disease progression and treatment response. This has been demonstrated across multiple domains: blood pressure variability over time better predicts cardiovascular outcomes compared to isolated readings; Prostate-Specific Antigen velocity provides improved mortality risk assessment versus single values in prostate cancer; and repeated biomarker measurements in Alzheimer's disease reveal how different markers change sequentially, with amyloid changes preceding metabolic decline and brain atrophy. Incorporating follow-up CT scans into deep learning models has been shown to significantly improve predictions of survival and disease progression in non-small cell lung cancer patients undergoing chemoradiation. Previous research has also demonstrated that integrated longitudinal omics profiling, including genomic, transcriptomic, proteomic, and metabolomic data, provides superior insight into dynamic health changes and disease progression compared to single-timepoint assessments. Serial measurements can identify dynamic changes in health status, potentially revealing early signs of treatment response or disease progression before they become clinically apparent. Recent multi-omics studies have further identified distinct personalized trajectories of aging, emphasizing the significance of individualized biomarkers for capturing variability in aging rates and associated disease risks. In this work, we introduce FAR as a biomarker calculated as the change in FaceAge between two time points divided by the time interval between the two facial photographs, providing a dynamic measure of aging over time. FAR value above 1 indicates accelerated aging (aging faster than expected), while FAR value below 1 suggests a decelerated aging. By analyzing FAR in a large cohort of cancer patients undergoing radiation therapy courses, we aim to establish whether this metric can serve as a prognostic marker for survival.

'You Haven't Aged a Day': Quantifying the Prognostic Value of Facial Aging

This transcript has been edited for clarity. Welcome to Impact Factor, your weekly dose of commentary on a new medical study. I'm Dr F. Perry Wilson from the Yale School of Medicine. I went to my 20 medical school reunion last weekend. Shout out to P&S Class of '06. Amazing to see all those folks again, and as you might expect, I heard (and said) quite a few times "my god, you haven't aged at all." But, of course, after 20 years this isn't technically true. I think what we probably mean when we say that is something along the lines of "you haven't aged nearly as fast as I would have expected." But for some reason, people don't react quite as well when you phrase it that way. It works the other way as well. We've all run into people, or patients, who look quite a bit older than we remember. It's generally considered rude to bring this up. But it turns out that we can actually quantify that subjective feeling of how old someone looks vs how old we expect them to be, and a new study suggests that metric might hold particular prognostic importance for cancer patients. There's no nice way to say it. Going through cancer treatment is incredibly stressful. There is the physical toll of chemotherapy, radiation, and the side effects thereof. There is the emotional toll of dealing with the diagnosis, reorganizing your life, worrying about the outcomes for you and your loved ones. It should not be surprising that cancer ages people. But that can be hard to quantify. Enter this paper, appearing in Nature Communications, that leverages AI to put hard numbers on those vague impressions of age we all have. The study enrolled 2276 patients with cancer undergoing radiation therapy. The majority, around 63%, had metastatic disease at baseline. As part of routine care, they had two photographs taken -- one at the start of therapy and one some time later. The researchers posed two questions. First, do people who appear older than their stated age at baseline do worse in the long term? Second, and I think of more interest, does the rate of aging have prognostic implications? First a note about how this all works. The authors used an open-source package called FaceAge, which is a custom software package trained on more than 40 million facial images. I was able to get a version of this running on Python on my computer to play around with. Here you see Barack Obama shortly before his inauguration in 2009. He was 47 at the time. According to FaceAge, he was 51.3. That's a difference of 4 years. In this study, they'd refer to that as the FaceAge deviation. On the right is Obama 7 years later, at age 54, towards the end of his presidency. According to FaceAge, he was 61.5, giving a FaceAge deviation of 7.5 years. We can construct the rate of aging by dividing his change in FaceAge by his change in time. Over 7 real years, his face aged 10.2 years, leading to a face aging rate of 1.45. This is the quantifiable age acceleration of the presidency. Let's move on to the patients in the study. What I'm showing you here is the FaceAge deviation at the first visit. This is how much older (or younger) than their stated age, the image model thinks they are. You can see a broad distribution, with some lucky individuals apparently looking 40 years younger than they actually should. On average though, the model seemed to be fairly well-calibrated, with the bulk of deviations from stated age sitting right around zero. It gets a bit more complicated when we look at the face aging rate. This is because the time between the two photographs wasn't standardized. Some people had photos as little as a week apart, a time in which the noise and error in the model would really dominate the results. This graph shows that effect. On the X axis, we have the time between the two photos, and on the Y axis, the face aging rate. To handle this issue, the authors stratify their analysis based on how far apart these photos were taken. Let's look at that analysis. We'll start with people whose photos were taken fairly close together, within a year of each other. Remember, over such a short time scale, there is going to be a lot of noise in the measurement, so the authors dichotomized the face aging rate as above or below 20. We're comparing people whose faces aged at least 20 times faster than expected to those whose faces did not age that fast. I'm shocked at how many patients were in that accelerated group -- about 20%. And they had a much higher all-cause mortality rate, 22% higher, in fact. For those who had a slightly longer time between pictures, ranging from 1 to 2 years, there was less noise in the measurement, allowing the researchers to dichotomize the data at a face aging rate of 10 times faster than expected. The death rate was 54% higher in this group. Among the 400 or so patients with more than 2 years between pictures, the authors simply divided the cohort at a face aging rate of 1 -- people aging faster than expected (like Obama) vs people aging slower than expected. Once again, a strong signal was apparent. Those aging faster than expected had a 60% higher mortality rate. These findings persisted after multivariable adjustment for things like sex, race, and cancer diagnosis. In fact, actual calendar age had no association with outcomes after accounting for FaceAge. I will note that there was no adjustment for Botox, fillers, or other nips and tucks that might affect how the face appears to an AI model. Critically, the findings also persisted after adjustment for baseline FaceAge deviation. This is best visualized in contour plots. What you see is that as baseline age deviation increases, the mortality risk increases; people who look older when they start therapy tend to do worse. But more than that, as you move up the plot, you see that for any given baseline age deviation, a faster rate of aging increases the risk still further. Analyses of the predictive power of these two factors actually found that the rate of face aging had more prognostic value than the baseline FaceAge. I wanted to talk about this study, because I like when things we don't usually quantify get quantified, but it's worth asking if there are any clinical implications here. At this point, I don't see us using AI to scan people's faces and make prognostic judgments. The oncologists will continue to use functional status metrics like ECOG to decide if ongoing chemotherapy has more benefit than risk. But what a study like this does do is open a door to asking a deeper biological question. What is going on under the surface that is leading to visible aging? Is it inflammation? Cortisol? Some issue in the lymphatic system? Malnutrition? In case you're tempted, as I was, to test out your own FaceAge, a number of sites will do this online. But a word of caution. Each model is different, and this study only tested the open source FaceAge model on GitHub. Just to show you how different these things can be, here's what the real FaceAge model thought of yours truly, current age 46. 52.9. That hurts. Sending the same photo to FaceAge.AI online, netted me a much nicer age estimate of 40. That model was kinder to President Obama as well, placing him at 47 at his first inauguration. So, yes, the devil is in the details. But it may turn out that the age you appear is more than skin deep, after all. F. Perry Wilson, MD, MSCE, is an associate professor of medicine and public health and director of Yale's Clinical and Translational Research Accelerator. His science communication work can be found in the Huffington Post, on NPR, and here on Medscape. He posts at @fperrywilsonand his book, How Medicine Works and When It Doesn't, is available now.

AI tool estimates biological age from photos to predict cancer outcomes

Mass General BrighamApr 28 2026 The Mass General Brigham research team behind FaceAge, an artificial intelligence (AI) tool that can estimate a person's biological age from a single photo, is reporting in a new study that estimating biological age from multiple photos taken over time can provide even more information about how well a person with cancer will do with treatment. Their results, published in Nature Communications, suggest that Face Aging Rate (FAR) - which uses photos to measure changes in biological age over time - can serve as a non-invasive biomarker for cancer prognosis. The new study analyzed two photographs from each of 2,279 patients with cancer, taken at different time points over the course of treatment. The researchers found that higher FAR was significantly associated with decreased survival probability. Deriving a Face Aging Rate from multiple, routine facial photographs allows for near real-time tracking of an individual's health. Our study suggests that measuring FaceAge over time may refine personalized treatment planning, improve patient counseling, and help guide the frequency and intensity of follow-up in oncology." Raymond Mak, MD, co-senior and corresponding author, radiation oncologist at Mass General Brigham Cancer Institute and faculty member in the Artificial Intelligence in Medicine (AIM) program, Mass General Brigham FaceAge is an AI tool that uses deep learning technologies to determine biological age from a person's face photo. In a study published last year, the investigators determined that patients with cancer were likely to appear about five years older than their chronological age per FaceAge, and that older FaceAge estimates correlated with worse survival outcomes after cancer treatment. In the new study, the researchers sought to learn what information FaceAge could provide when applied to multiple photos of the same person taken over time. They inspected facial photos from a cohort of patients with varying types of cancer who received at least two courses of radiation therapy at Brigham and Women's Hospital between 2012 and 2023. The photos were taken as part of the routine clinical workflow at each separate radiation therapy course. FAR was calculated as the change in FaceAge between these two time points, divided by the time interval. The researchers also calculated FaceAge Deviation (FAD), which estimates how biologically old or young the patient looked in a single face photo relative to chronological age. Median FAR results indicated that patients' facial aging outpaced their chronological aging by 40%. Higher FAR, or accelerated aging, was associated with lower survival, and the effect was strongest when the interval between photos was two years or more. Additionally, patients with both high FAD and FAR values were significantly more likely to have poorer survival probabilities. However, FAR was more likely to predict survival outcomes stably over longer intervals than FAD - indicating that dynamic measurements might be more reliable than single timepoint readings. The authors suggest that integrating FAR with baseline FAD could provide a more nuanced and informative measure of an individual's evolving health status. Further research is needed to evaluate FaceAge and FAR in more diverse populations. "Tracking FaceAge over time from simple photos offers a non-invasive, cost-effective biomarker with potential to inform individuals of their health," said study co-author Hugo Aerts, PhD, director of the AIM program at Mass General Brigham. "We hope with continued study we can learn how FaceAge may provide prognostic information for patients with other chronic diseases and for healthy individuals." This research provides additional demonstrations of the potential clinical utility of FaceAge. In another recent study published in JNCI: Journal of the National Cancer Institute, FaceAge was tested on more than 24,500 cancer patients over the age of 60 who received radiation therapy. FaceAge was older than chronological age in 65% of the patients. Those with a FaceAge estimate 10 or more years older than their chronological age had significantly worse survival outcomes, while those with an estimate of five or fewer years had better outcomes. Ongoing and future studies, including prospective trials, will continue to investigate FaceAge outcomes in patients with different cancers and other diseases. The research team has now also launched an institutional review board-approved web portal that allows the general public to submit their own face photographs to get a FaceAge assessment and participate in research to advance this technology at faceage.bwh.harvard.edu. Mass General Brigham Journal reference: Haugg, F., et al. (2026). Face aging rate quantifies change in biological age to predict cancer outcomes. Nature Communications. DOI: 10.1038/s41467-025-66758-w. https://www.nature.com/articles/s41467-025-66758-w