Devi Sridhar: AI cancer survival models trained on 40 million faces outrun their evidence

A machine that estimates how long a cancer patient has left by reading a selfie sounds like medicine's moonshot. Devi Sridhar, professor of public health at Edinburgh, is not convinced — and the details of the FAHR-Face study, published this month in The Lancet Digital Health, explain why.

Researchers at Mass General Brigham trained their model on more than 40 million facial images. One component, FAHR-FaceAge, estimates biological age from the forehead, the wrinkles between nose and mouth, and the temples. A second, FAHR-FaceSurvival, was trained on almost 35,000 photos of cancer patients and focused on the area below the eyes and the bridge of the nose — features Hippocrates flagged in 400BC as signs of wasting and approaching death. The researchers found the AI-estimated facial age was significantly associated with survival in 18 of 28 cancer categories.

Sridhar's concern is less the technology than the comparison being sold. The headline claim — that AI predicts cancer survival better than a doctor — rests on benchmarks where the machine is measured against clinicians working under conditions no model shares: time pressure, incomplete records, and patients who do not arrive as clean data points. Prof Miroslaw Bober, quoted in the piece, concedes the tool needs further testing against doctors' own judgments and across different hospitals and patient groups before anyone declares victory.

The results also carry a warning the press releases omit. After adjustment for other factors, the facial-age measure was not significantly associated with survival for Black patients. The same pattern has shown up in AI models analyzing cardiac magnetic resonance scans used to diagnose and monitor heart disease. A tool trained on faces that skews pale will read darker faces badly — and in oncology, a bad read is not an academic problem.

The deeper issue is structural. Hospitals and AI vendors both profit from the claim that software outperforms physicians: one sells platforms, the other sells adoption. Nobody in that chain bears the cost of a patient triaged wrongly by an overconfident model. The clinician who overrides the algorithm risks career and liability; the vendor faces neither.

Sridhar does not dispute that facial features track health — Hippocrates got there first. What she disputes is the leap from a Lancet paper to a product. Until the tool survives testing against actual doctors in actual hospitals, the selfie remains what it always was: a photograph.