MS-ECG-FM: Towards a More Universal Electrocardiogram Foundation Model for Health Monitoring using Multi-source Contrastive Learning

Robert A. Lewis, I-Min Chiu, Kyle Verrier, Karthik Jayaraman Raghuram, Francoise Marvel, Salar Abbaspourazad, Anshuman Mishra, Guillermo Sapiro, Andrew C. Miller†, Joseph Futoma† (Apple; MIT; Johns Hopkins; Princeton)

Summary

ECG transformer with lead-level tokens, contrastively aligned to five clinical note types linked to each MIMIC-IV-ECG recording (machine ECG reports 800,034; cardiologist ECG reads 490,780; echo 81,821; chest X-ray 715,465; discharge summaries 331,793). Echo/X-ray are matched to the nearest report within 7 days; discharge summaries by visit. Linear-probe AUROC versus MERL, D-BETA, MELP, ECGFounder and a random encoder, with paired patient-cluster bootstrap significance tests.

Reported benchmarks