Machine-learning model detects within-person facial changes in acromegaly

Researchers conducted a multicenter, international cohort study using retrospective photographs from people with acromegaly and other pituitary tumors. Of 582 participants, 110 had usable frontal photographs and 69 had photographs from two timepoints: 50 people with acromegaly and 19 tumor controls. A face-recognition model and a general-image model were used to score facial patterns; photographs were paired by reported disease stage when available, or chronologically otherwise.

The face-recognition score distinguished held-out faces with advanced acromegaly from control faces, with an area under the curve of 0.858. Among people with acromegaly, 76.0% shifted in the predicted direction between photographs (median within-person shift 0.59; p=0.0003); the proportion was 81.5% in clean, untreated pairs (p=0.0008). The general-image model showed no significant within-person change. Pre-diagnosis untreated photographs showed modest separation from control baselines (AUC 0.65).

The findings suggest that facial change associated with acromegaly may be measurable in routine photographs, but do not establish a clinically validated monitoring or screening tool. The analysis was retrospective, and only 69 participants had paired images; the authors call for further evaluation.