#26 · ML / AI methods

Smartphone video and images yield 3D kinematics and body composition phenotypes

The dominant thread is extracting quantitative human phenotypes from consumer smartphone capture rather than laboratory instrumentation. OpenCap Monocular recovers 3D skeletal kinematics and kinetics from a single smartphone video through pose refinement and constrained skeletal modeling, targeting musculoskeletal dynamics. The gait foundation model operates on 3D skeletal motion from 3,414 adults, learning embeddings that predict age, BMI, visceral fat and several multi-system phenotypes. PhotoScan applies a deep learning model pretrained on UK Biobank and fine-tuned on clinical cohorts to smartphone images, estimating body composition as an alternative to BMI for cardiometabolic risk assessment. What is new across these preprints is the pairing of monocular 3D reconstruction with representation learning and body-composition inference from routine devices. The evidence here is limited to preprint abstracts, so validation scope and comparative performance remain to be established from full texts.