Machine Learning Quantifies Within-Subject Facial Change in Acromegaly.

OBJECTIVE: Machine-learning models can distinguish acromegaly from controls in single photographs, but whether gradual facial change can be measured within the same person from ordinary photographs is unknown. We tested whether a person's face shifts measurably toward an acromegaly-associated pattern over time. DESIGN: Multicenter, international, cross-sectional cohort study with within-subject longitudinal analysis. METHODS: People with acromegaly and other pituitary tumors contributed retrospective photographs. Two timepoints per person were paired by self-reported disease stage when available, with chronological pairing otherwise. A face-recognition model and a general-image model described each image, and leave-one-patient-out scoring quantified within-person movement toward the acromegaly-associated pattern. RESULTS: Of 582 participants, 110 contributed usable frontal photographs and 69 had two timepoints (50 acromegaly, 19 tumor controls). The facial score separated held-out advanced-disease from control faces (AUC 0.858). In acromegaly, the median within-person shift was 0.59 and 76.0% moved in the predicted direction (p=0.0003); 81.5% of clean untreated pairs moved as predicted (p=0.0008). The three contrast strata ordered directionally, although the trend was not significant (p=0.15). The 15 non-Cushing aging controls showed no significant shift (p=0.72), and the between-group difference remained after adjustment for photograph interval (p=0.029). Pre-diagnosis untreated baselines showed modest separation from all control baselines (AUC 0.65, one-sided p=0.033). The general-image model showed no within-person change (p=0.32). CONCLUSIONS: Acromegaly-associated facial change is measurable within individuals from ordinary photographs, supporting further evaluation of longitudinal photograph-based monitoring and earlier recognition.