Summary
CSFM (Cardiac Signal Foundation Model) is a transformer-based foundation model pretrained on 1.7 million individuals via generative masked pretraining across both temporal and channel dimensions. Published in Nature Machine Intelligence (Feb 24, 2026, open access). Key innovation: device-agnostic — one model handles 12-lead clinical ECGs, single-lead ECGs (wearable-grade), and PPG, including mixed ECG+PPG inputs.