Researchers at Mass General Brigham have developed artificial intelligence models that analyze facial images to estimate biological age and forecast cancer survival, pushing AI cancer survival prediction into territory once occupied only by a doctor’s trained eye. The work, presented by study author Raymond Mak and colleagues, rests on two models known collectively as FAHR-Face.
Two Models, Two Purposes
FAHR-FaceAge estimates biological age from facial features. Its training set was vast: more than 40 million facial images. The model draws on regions that people commonly associate with aging, including the forehead, nasolabial wrinkles and the temples.
FAHR-FaceSurvival serves a different purpose. It was trained on nearly 35,000 photographs of cancer patients and predicts survival outcomes. Rather than scanning the whole face, it concentrates on the area below the eyes and across the nose bridge.
Mak said those facial areas are linked with dying. The observation is not new in medicine, even if the computational approach is.
An Idea Older Than Modern Medicine
Hippocrates, writing around 400 BC, described facial signs that could signal a risk of death. His list included a sharp nose, sunken eyes and fallen temples. Mak noted the historical parallel, and the continuity is striking: the regions FAHR-FaceSurvival prioritizes overlap with features physicians have observed for millennia.
Doctors have long incorporated facial appearance into clinical assessment. A clinician may note whether a patient looks older or younger than the stated age, using that impression as one thread in a broader evaluation. The FAHR-Face models attempt to formalize that habit at scale.
Skepticism About Individual Predictions
Not everyone is convinced the models can deliver on their most ambitious promise. Devi Sridhar, chair of global public health at the University of Edinburgh, expressed doubts about predicting individual survival. Her concern centers on whether an AI system trained on photographs can reliably forecast what will happen to a single patient.
Sridhar’s clinical colleague found the development interesting, according to the brief. The split reaction captures a broader tension in medical AI: aggregate pattern recognition can be impressive while individual-level prediction remains uncertain.
What Remains Unanswered
The research leaves several questions open. Accuracy rates for the models were not provided. Peer review status is not stated. The demographic composition of the training data is absent, and clinical validation details are not described. Without those elements, it is difficult to judge how the models would perform outside the datasets on which they were built.
The underlying premise, however, is straightforward. AI is improving at spotting patterns in medical data, and facial images are a rich source of visible signals. Whether those signals can be translated into dependable survival forecasts for individual cancer patients is a question the FAHR-Face work raises without fully settling.
For now, the models represent an attempt to bridge ancient clinical observation and modern machine learning. Hippocrates could only describe what he saw. The FAHR-Face researchers are testing whether a machine can see the same signs and quantify them.
