Seeing beyond BMI: Estimating cardiometabolic risk with smartphone imagery

PhotoScan, a deep-learning model trained on 35 k UK Biobank participants and fine-tuned with 677 adults, predicts three-dimensional body composition from standard 2D smartphone images. In cross-validation it achieved 2.15% mean absolute error for body-fat percentage, 0.107 for Android-to-Gynoid ratio, and 0.094 for Visceral-to-Subcutaneous ratio, outperforming smartwatch BIA (2.91% MAE). Independent validation yielded similar errors (2.13% BF%, 0.085 for both ratios) and allowed insulin-resistance prediction comparable to DXA.