ECG-LLM reads 679K ECGs, matches specialists, detects hidden cardiac phenotypes

ECG-LLM is an ECG-conditioned large language model developed by researchers at TUM and Klinikum rechts der Isar. It was trained on four cohorts totaling 679,112 ECG studies from 186,409 patients, using a novel multimodal-to-language supervision approach that converts ECG signals, clinical context, and imaging data (CMR, ECHO) into clinically structured question-answer pairs. This allows the model to answer diverse cardiovascular questions from a 12-lead ECG alone, going beyond fixed-label prediction to support patient-specific reasoning.

The model strongly predicts CMR-derived phenotypes like ventricular/atrial volumes and function, and detects critical echocardiographic signs such as increased LV wall thickness, aortic stenosis, and right-ventricular systolic dysfunction. On standard ECG understanding benchmarks (diagnostic report generation and ECG-QA), ECG-LLM matches or exceeds existing baselines. By enabling question-driven cardiovascular reasoning, it could help general practitioners triage patients when specialist review is delayed.