Google Research's PhotoScan Predicts Insulin Resistance at 0.760 AUROC From Smartphone Photos
Updated
Updated · Google Research · Aug 17
Google Research's PhotoScan Predicts Insulin Resistance at 0.760 AUROC From Smartphone Photos
1 articles · Updated · Google Research · Aug 17
Summary
Google Research said its PhotoScan model estimated body composition from standard smartphone photos and predicted insulin resistance with a 0.760 AUROC, close to DXA-based models at 0.773.
35,000-plus UK Biobank records and a 677-adult fine-tuning cohort trained the system to infer body fat, android-to-gynoid ratio and visceral-to-subcutaneous fat ratio—metrics linked more closely to insulin resistance than BMI alone.
2 independent cohorts showed consistent body-composition accuracy: body-fat mean absolute error was 2.15 in the PhotoBIA cohort and 2.13 in MetabolicMosaic, while A/G and V/S errors stayed near 0.085-0.107.
Smartwatch BIA underperformed because it mainly estimates body-fat percentage; in Google’s tests, adding BIA to demographics did not improve insulin-resistance classification.
The project remains a research prototype, but Google framed it as a scalable, non-invasive screening path that could later be combined with wearables, glucose data and blood biomarkers.