Facial foundation model estimates clinical biomarkers

In human datasets, researchers developed MedicalFaceFound, a facial foundation model trained on over 10 million images to predict clinical biomarkers non-invasively. The team evaluated the model across 62 biomarkers representing eight physiological systems.

MedicalFaceFound outperformed Swin-Large and ResNet18 models on 45 of 62 (73%) biomarkers. In tests across four independent external cohorts, it achieved an overall median Pearson's r of 0.172, showing its highest correlations for red blood cell count (median r = 0.478), eGFR (median r = 0.442), and HDL-C (median r = 0.430). The model also surpassed polygenic risk score models for 14 of 26 biomarkers, and face-estimated cardiovascular metrics were associated with coronary stenosis with an AUC of 0.66. The architecture retained predictive accuracy with 400 labeled samples and was successfully deployed as a smartphone application.