Seeing beyond BMI: Estimating cardiometabolic risk with smartphone imagery

Details

Google Research has published PhotoScan, an investigational deep learning framework that estimates 3D body composition metrics — body fat percentage, android-to-gynoid fat ratio, and visceral-to-subcutaneous fat ratio — directly from 2D smartphone photos. Pre-trained on over 35,000 UK Biobank participant records and fine-tuned on 677 diverse adults, the model targets HOMA-IR prediction (score >2.9 indicates insulin resistance).\n\nThe clinical gap PhotoScan addresses is real. DXA scans, the gold standard for body composition, are expensive, radiation-emitting, and require specialized infrastructure. Insulin resistance precedes type 2 diabetes by years and remains widely underdiagnosed. A smartphone-based screen could shift early metabolic disease detection into everyday consumer contexts.\n\nThe claimed accuracy — comparable to DXA scans in a clinical research setting — is significant if it holds up at population scale. The model captures compositional biomarkers (A/G ratio, V/S ratio) that outperform BMI as metabolic predictors. No privacy architecture, compute requirements, or deployment timeline are disclosed. This remains investigational research, not a product announcement.\n\nThe approach extends a growing trend of passive smartphone health monitoring. If PhotoScan generalizes across diverse demographics, it could enable large-scale metabolic screening without clinical equipment — though broader demographic validation beyond the 677-person cohort is needed.