ResearchPapers, evidence & method
Devi Sridhar questions Mass General Brigham's FAHR-Face model, which predicts cancer survival from facial photos
Mass General Brigham's FAHR-Face, a foundation model pretrained on over 40 million face photos, separates cancer patients by mortality risk in retrospective radiotherapy data. Devi Sridhar questions whether it adds to clinical judgement, citing uneven subgroup results, selfie variability and consent.
From the The Guardian · AI article of October 23, 2026 and cited sources: In the authors' own validation figures, patients in the highest-risk quartile died at roughly three times the rate of the lowest quartile (adjusted hazard ratio about 3.2). [5] [3] [2]
SOURCES[5] arXiv 2506.14909 – FAHR-Face preprint arxiv.org[3] News-Medical – technical summary news-medical.net[2] Science Media Centre – expert reaction sciencemediacentre.org
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- In the authors' own validation figures, patients in the highest-risk quartile died at roughly three times the rate of the lowest quartile (adjusted hazard ratio about 3.2). [5] [3] [2]
- Original source
- The Guardian · AI article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- Studies
- arXiv 2506.14909 (Jun 2025) [5] date from arXiv ID
- When
- October 23, 2026 article published
- Covered
- Oct 23 edition
- Evidence
- 8 sources across 8 sites
- 1Can AI really predict cancer survival better than a…theguardian.com
- 2Science Media Centre – expert reactionsciencemediacentre.org
- 3News-Medical – technical summarynews-medical.net
- 4PubMed – FaceAge 2025pubmed.ncbi.nlm.nih.gov
- 5arXiv 2506.14909 – FAHR-Face preprintarxiv.org
- 6EurekAlert – Mass General Brigham press releaseeurekalert.org
QuietHow it’s measured
Hacker News · 2 likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
In the authors' own validation figures, patients in the highest-risk quartile died at roughly three times the rate of the lowest quartile (adjusted hazard ratio about 3.2). Of the patients flagged as highest-risk, 52.5% died within 90 days, compared with about 2% of the lowest-risk group. These results rank patients against one another. They do not show that the model gives accurate survival probabilities for an individual patient. [5] [3] [2]
FAHR-Face has not been tested against clinicians' judgement or at other hospitals. Its 2025 predecessor, FaceAge, is the only version in the reviewed sources that was tested alongside doctors: adding it raised physicians' accuracy (AUC) at predicting survival in palliative patients from 0.74 to 0.80. Sridhar asks whether the new model adds anything beyond the rough 'looks older than their age' judgement doctors already make. [2] [4] [1]
Claims that the model works equally well for everyone and copes with varied photos are disputed. The developers say it generalises across demographic groups and holds up across pose and lighting. Other reviewers report that the facial-age signal was not significant for Black patients after adjustment, that risk separation was weaker for Black and Hispanic patients, and that smiling or a lighting change shifted age estimates by about two years. Botox, fillers and GLP-1 drugs also change the facial regions the model relies on. [5] [2] [3] [6] [1]
The evidence shows the model can rank patients by risk in past data. It does not show the model improves care. Health systems should treat it as a research tool, and selfie-based health apps face the same bias and consent questions.
Executive brief
The Mass General Brigham model behind the "selfie predicts cancer survival" headlines has never been tested against doctors' judgement, according to an independent expert reviewing it (Science Media Centre). FAHR-Face was pretrained on more than 40 million face photos. In its validation data, 52.5% of the patients it flagged as highest-risk died within 90 days, against about 2% of the lowest-risk group (News-Medical). All of this comes from retrospective data on radiotherapy patients. In the Guardian, Devi Sridhar asks whether it adds anything to what clinicians already do. She also points to uneven results across racial groups, how much selfies vary, and the fact that the public website enrols uploaders as research participants.
What changed and event timeline
FaceAge precursor published
An earlier Lancet Digital Health model estimated "facial age" from photos of 6,196 cancer patients. Older-looking faces were linked to worse survival, and adding the model improved physicians' palliative-care survival predictions ().
FAHR-Face preprint
Haugg, Aerts, Mak and colleagues posted the foundation-model version on arXiv. It reported an adjusted hazard ratio of 3.22 between the highest and lowest risk quartiles ().
Peer-reviewed paper and expert reaction
The press release accompanied publication in The Lancet Digital Health (). Miroslaw Bober called it a promising proof of concept that is not ready for clinical use ().
Trade coverage spreads the numbers
Oncology and health outlets repeated the hazard ratios and 90-day mortality figures (;).
Sridhar's critique
The Guardian column questions whether the model is useful in practice, and raises concerns about selfie reliability and how the public website handles consent ().
Capabilities and access
- FAHR-Face: a ViT masked-autoencoder foundation model. Its weights are public under CC BY-NC-SA 4.0 (non-commercial) and the code under Apache 2.0 (Hugging Face).
- FAHR-FaceAge: estimates biological age from a face photo. Weights are released (Hugging Face).
- Weights are not public (Hugging Face).
Read the full section
- FAHR-Face: a ViT masked-autoencoder foundation model. Its weights are public under CC BY-NC-SA 4.0 (non-commercial) and the code under Apache 2.0 (Hugging Face).
- FAHR-FaceAge: estimates biological age from a face photo. Weights are released (Hugging Face).
- FAHR-FaceSurvival: predicts survival risk. Weights are not public (Hugging Face).
- Public website: faceage.bwh.harvard.edu returns an age estimate, and uploaded photos become study data (EurekAlert).
Technical analysis for researchers and developers
- Pretraining: self-supervised masked autoencoding with 75% of image patches hidden, on more than 40 million consumer images (arXiv).
- FaceAge fine-tuning: two stages, balanced by age, on ten public datasets. The model's age estimates are off by 5.1 years on average (arXiv).
- Survival fine-tuning: on about 34,000 photos of cancer patients. The preprint gives 34,389, while later coverage says 34,363 (arXiv; ASCO Post).
Read the full section
- Pretraining: self-supervised masked autoencoding with 75% of image patches hidden, on more than 40 million consumer images (arXiv).
- FaceAge fine-tuning: two stages, balanced by age, on ten public datasets. The model's age estimates are off by 5.1 years on average (arXiv).
- Survival fine-tuning: on about 34,000 photos of cancer patients. The preprint gives 34,389, while later coverage says 34,363 (arXiv; ASCO Post).
- Validation cohorts: 38,183 patients at Harvard radiotherapy centres and 4,892 in the Netherlands for the age model; 3,820 patients for the survival model (News-Medical).
- What the evidence shows: the results are hazard ratios and risk-group comparisons, so they show the model ranks patients against each other. It has not been shown to give accurate survival probabilities for an individual patient (SMC).
Claims and evidence
- Risk separation (authors' figures): the highest-risk quartile died at more than three times the rate of the lowest (adjusted HR 3.22–3.23).
- Fairness across groups, where the sources disagree: the authors report the models generalise across age, sex, race and cancer type (arXiv).
- Robustness to photo conditions, also disputed: the authors report stable predictions across pose, lighting, makeup and most plastic surgery (EurekAlert).
Read the full section
- Risk separation (authors' figures): the highest-risk quartile died at more than three times the rate of the lowest (adjusted HR 3.22–3.23). Each 0.1-point rise in the survival score came with 17% higher mortality hazard (arXiv; News-Medical).
- Fairness across groups, where the sources disagree: the authors report the models generalise across age, sex, race and cancer type (arXiv). Bober notes the facial-age measure was not significant for Black patients after adjustment (SMC). News-Medical reports weaker stratification for Black and Hispanic patients (News-Medical).
- Robustness to photo conditions, also disputed: the authors report stable predictions across pose, lighting, makeup and most plastic surgery (EurekAlert). Bober reports that smiling or a lighting change shifted age estimates by about two years on average (SMC).
- Results by cancer type: the facial-age signal was significant in 18 of 28 cancer categories (Guardian).
Context and prior work
- Clinical tradition: doctors have long written "appears older than stated age" in notes as a rough gauge of health (Guardian).
- FaceAge (2025): the earlier model is the only version in the reviewed sources tested alongside clinicians.
- What's new in FAHR-Face: the authors say it outperformed FaceAge at predicting survival (arXiv).
Read the full section
- Clinical tradition: doctors have long written "appears older than stated age" in notes as a rough gauge of health (Guardian).
- FaceAge (2025): the earlier model is the only version in the reviewed sources tested alongside clinicians. Adding it raised physicians' AUC for predicting survival from 0.74 to 0.80 in palliative patients (PubMed).
- What's new in FAHR-Face: the authors say it outperformed FaceAge at predicting survival (arXiv). The main change is the foundation-model design: one pretrained model can be fine-tuned into several facial biomarkers instead of training each from scratch (EurekAlert).
Limitations, safety and contested findings
- Study design: retrospective, radiotherapy-only cohorts with limited demographic diversity and no prospective validation (News-Medical).
- No comparison with doctors: the model has not been tested against clinicians' judgement or across other hospitals (SMC).
- Changes over time: claims that it can track ageing or lifestyle changes over time are premature, because this was not tested (SMC).
Read the full section
- Study design: retrospective, radiotherapy-only cohorts with limited demographic diversity and no prospective validation (News-Medical).
- No comparison with doctors: the model has not been tested against clinicians' judgement or across other hospitals (SMC).
- Changes over time: claims that it can track ageing or lifestyle changes over time are premature, because this was not tested (SMC).
- Faces change: Botox, fillers and GLP-1 drugs alter the facial regions the model relies on (Guardian).
- Consent and privacy: people who upload a selfie to the public site become research participants (Guardian).
- Conflicts of interest: Bober declares roles as CTO of Forecom.AI and as an AI consultant to Pfizer, BMS and the NHS (SMC).
Business and practitioner implications
- Commercial use: the non-commercial license rules out building products on the released weights, and the survival model is not released at all (Hugging Face).
- For researchers: the released backbone makes it practical to fine-tune new facial biomarkers.
- For health systems: at this stage the model is a research-grade aid for triage, not a prognostic device.
Read the full section
- Commercial use: the non-commercial license rules out building products on the released weights, and the survival model is not released at all (Hugging Face).
- For researchers: the released backbone makes it practical to fine-tune new facial biomarkers. Any new claim should come with results by subgroup and by photo condition, given the disputes above.
- For health systems: at this stage the model is a research-grade aid for triage, not a prognostic device. The authors themselves describe it as supporting clinical judgement, not replacing it (EurekAlert).
- For consumer apps: any app that reports health information from a selfie inherits these consent and bias questions.
The source trail.
Sources (8)
Can AI really predict cancer survival better than a doctor? I have my doubts | Devi Sridhar
Article text retrieved; extracted text may omit tables or interactive elements.
theguardian.com