One Photograph, Several Readings: Checking an AI Prediction of Cancer Patient Survival Time from a Face Photograph Against the Camera and Software's Own Settings

A foundation model trained on over 40 million face photographs can now estimate biological age and predict the survival of cancer patients from a single photograph. Today the photograph is taken once and the model gives one survival score, with no sign of how firm that score is. This concept note describes a simple check that can sit on top of such a model. From the photograph that is already stored, the computer makes two to four versions by changing only the photograph's own processing settings, within the range the camera, the capture software and the clinic permit. The same model reads every version, unchanged. If the survival scores stay together, the prediction is firm. If they spread out, it depends on a setting. One number, the spread value, shows which case it is. A computer-decided shunting action then moves the case that needs care to the front of the clinician's list while all other cases carry on. The check adds no device, no visit and no change to the model, and the decision stays with the clinician. The note explains the method in plain steps, shows how it speaks to the lighting and image-quality limitation named by the authors of the published model, and compares it with earlier work and patents.