On October 19, a Mass General Brigham group published FAHR-Face in Lancet Digital Health (Haugg, Lee and colleagues). It is a set of models that estimate apparent age and survival risk from a single photograph of a cancer patient's face. Scientific American covered it the same day under the headline "Can AI predict whether you'll die from cancer from just an image of your face?"

This review compares the article with the paper, assesses the study design, and notes the sources of bias.

Three models, stacked. A foundation model learned general facial structure from more than 40 million unlabeled internet photographs (the WebFace42M dataset) by hiding patches of each image and reconstructing them. Two task models were then fine-tuned on top of it.

FAHR-FaceAge estimates apparent age. It was tuned on 743,152 age-labeled public photos (IMDB-WIKI, MORPH, UTKFace and others) and reaches a mean absolute error of 5.1 years on a held-out public set. Its clinical output is FAD, the gap between face age and calendar age.

FAHR-FaceSurvival produces a 0 to 1 risk score. It was trained on photos of 34,363 patients treated with radiotherapy at six Harvard-affiliated sites from 2008 to 2023, tested on a random 10 percent holdout of 3,820, and checked externally on 4,892 radiotherapy patients in Maastricht. The photos are the identification pictures taken during routine radiation oncology care, not selfies.

What holds up. The cohort is large, the external validation is in a different country, the survival analysis adjusts for cancer site, treatment intent, year and radiation technique, reporting follows TRIPOD+AI, and the code is public. Early mortality rises across all ten risk deciles.

No clinician comparator. The covariates leave out performance status (ECOG or Karnofsky), weight loss, albumin and stage. These are the standard measures of the clinical impression Mak tells Scientific American the model is meant to quantify. The group's own 2025 FaceAge paper asked 10 clinicians to predict short-term survival from photos of 100 palliative patients and found them only slightly better than chance, improving when given FaceAge. This paper does not repeat that comparison, so the model's added value over clinician judgment is not measured.

Adjustment does most of the work. The survival model's top-versus-bottom quartile hazard ratio falls from 8.75 to 3.23 once cancer site and treatment intent enter. Per 0.1 of score, it falls from 1.32 to 1.17. FAD falls from 1.30 to 1.14 per decade. Much of the unadjusted association is explained by information already in the medical record. The adjusted association remains significant.

The age model drifts. In Maastricht, FaceAge overestimated age by 4.7 years on average and the median FAD was 4.4 years, against 1.1 at Harvard. The paper does not explain the difference. Camera, lighting and population differences are all possible causes. Site-level recalibration would be needed before individual scores could be interpreted.

Hazard is not mortality. The abstract says the highest-risk quartile had "more than triple the mortality" of the lowest, citing an adjusted hazard ratio of 3.23. A hazard ratio is an instantaneous rate comparison. The five-year Kaplan-Meier survival figures are 78.1 percent and 17.7 percent.

Shortcuts were not tested. These are clinical photos taken in a radiation department. The stress tests covered make-up, surgery, pose and lighting on public datasets. They did not cover the things that appear on sick faces in hospitals: nasal cannula tubing, hair loss, steroid-related facial swelling, visible head and neck tumors. The survival model's attention falls below the eyes and across the nose, where oxygen tubing also runs. The paper does not test whether such artifacts contribute.

The prediction can cause itself. This is retrospective data. A patient who looked very sick may have been offered less aggressive treatment, and then died sooner. A model trained on those outcomes cannot separate appearance from the clinical response to it.

Demographic. The abstract claims generalizability across race. The Harvard training population was 85 percent White and 1.6 percent Hispanic. In the survival test set, the Hispanic subgroup was 59 people and the separation between high and low scores was not significant (p=0.15). The authors' own subgroup analysis found weaker face-age risk stratification in Black and Hispanic patients than in White and Asian patients. Subgroups with small numbers do not support the generalizability claim.

The reference faces. The age model learned what a given age looks like partly from IMDB-WIKI and the Cross-Age Celebrity Dataset, which are photographs of actors and public figures. These images are not representative of the general population's appearance at a given age.

Consent. The 40 million foundation images were collected from the internet. The Harvard patients' photos were used under an IRB waiver of consent, justified as minimal risk, retrospective, and involving a cohort in which many patients had died. Both are standard practice, and neither involved individual consent.

Financial and authorship. Mass General Brigham has filed a US provisional patent on next-generation facial health algorithms (63/769,840), disclosed by Haugg, Mak and Aerts. Funding was a European Research Council grant. The benchmark the new model beats is the group's own previous model.

The article as intake form. In a parenthetical, Scientific American invites readers to help train the algorithm and receive a face age estimate by sharing their own photographs through a link. The article does not describe the consent or data-use terms for those photos.

Value. Clinicians already estimate prognosis from a patient's appearance, and those estimates vary between clinicians and are not recorded in a standard form. A photograph is inexpensive, noninvasive and already collected in radiation oncology. A validated score could make that assessment consistent and auditable, and could flag frailty where a formal assessment is impractical.

Comparator. A score is useful when it improves a decision. Without a comparison against performance status or clinician estimates, it is not known whether FaceSurvival adds information or restates what clinicians already see.

Transportability. The 4.7-year shift in the Maastricht cohort shows calibration can change between sites. Deployed models encounter cameras and populations beyond those studied.

Feedback. Once a score is in the chart, it can influence treatment, which affects survival, which becomes training data for later versions.

Secondary use. Face photographs are widely available outside clinical settings. Liao raises insurer and employer use in the article. The paper studies radiotherapy identification photos only.