- Published on
Can AI really predict cancer survival better than a doctor? I have my doubts | Devi Sridhar
- Authors
- Name
- DeepHorus
New headlines suggested that a single selfie could predict how long a cancer patient might live, after researchers at Mass General Brigham unveiled an AI model called FAHR-Face. The tool, described in The Lancet Digital Health, was trained on more than 40 million facial images to estimate biological age and predict survival. Devi Sridhar, writing in her column, argues the development is intriguing but far from conclusive. The claim matters because it promises a cheap, noninvasive way to flag risk, while raising hard questions about whether AI can truly outperform a doctorâs judgment or simply dress up an old bedside observation.
Doctors have long used facial appearance as a rough proxy for health, writing notes such as âappears older than stated age.â Hippocrates described the ânose sharp, the eyes sunken, the temples fallen inâ as signs of muscle wasting and mortality risk. The new work formalizes that intuition with deep learning. FAHR-FaceAge estimates age from the forehead, the wrinkles between the nose and mouth, and the temples. FAHR-FaceSurvival, trained on almost 35,000 photos of cancer patients, focuses more closely on the area below the eyes and the bridge of the nose, regions the authors link with dying.
The published results are mixed. AI-estimated facial age was significantly associated with survival in 18 of 28 cancer categories, meaning it worked for some cancers but not others. Dr Raymond Mak, one of the studyâs authors, connected the facial regions to Hippocratesâ ancient descriptions. Prof Miroslaw Bober cautioned that the tool still needs testing against doctorsâ own judgments and across different hospitals and patient groups. Notably, the facial-age measure was not significantly associated with survival for Black patients after adjustment for other factors, adding to concerns about racial bias in medical AI.
Technically, these models are built on retrospective data and statistical associations across thousands of patients. Deep neural networks learn pixel patterns that correlate with age or survival in the training set, but correlation is not the same as individual prediction or clinical utility. The input is also fragile. Lighting, camera angle, facial expression, time of day, Botox, fillers, facelifts and GLP-1 drugs can all alter facial fat and shape. The researchers say they tested their age model against lighting, expression and cosmetic surgery, but if the predictions rest on visible signs of aging, some error seems unavoidable.
In the short term, a validated version could augment triage or prompt closer clinical review, but only if it beats or complements existing tools. In the long term, as anti-ageing treatments become more common, visible age and underlying health may diverge, weakening the modelâs assumptions. Deployment without diverse validation could also widen disparities. Because the study is retrospective, prospective trials and direct comparisons with clinician predictions are needed before anyone uses a selfie to guide cancer care.
The reaction reflects that gap between promise and evidence. Clinicians find the research interesting because it quantifies a skill doctors already use, but skeptics warn that headlines have outpaced the data. Boberâs comments speak to broader concerns about reproducibility, subgroup performance and bias. Patient advocates may worry that facial predictions could be used to set expectations or deny care. The research team is impressive, but the jump from population-level association to a reliable prognosis for one patient remains large.
Looking ahead, the field needs prospective studies across diverse hospitals, comparisons with doctorsâ judgments and robust subgroup analyses. Regulators will eventually need standards for facial biomarkers and bias auditing. For now, a selfie is not a crystal ball; it is a research signal. The promise is real, but unproven, and clinical adoption should wait for rigorous validation rather than headline-driven enthusiasm.
