Predicting cardiovascular disease risk using photoplethysmography and deep learning

Summary

Google Research's Health AI team demonstrates that deep learning applied to fingertip PPG signals from the UK Biobank, combined with basic demographic data (age, sex, smoking status), can predict 10-year risk of major adverse cardiovascular events (MACE) non-inferiorly to blood pressure screening — without requiring physical examination or laboratory tests. The model generates PPG-based embeddings using a CNN encoder, then combines these with demographics to produce a calibrated cardiovascular risk score. The key finding is that the deep learning system (DLS) matches the specificity/sensitivity trade-off of blood pressure screening (a standard clinical reference) at matched specificity of 63.6%.