A new AI framework teaches models to read heartbeats the way cardiologists do: patch by patch, not just at a glance.
Researchers built a system called FOCAL that pairs small segments of an ECG waveform with specific words in a clinical report, using a mathematical technique called optimal transport to match the two. Most existing ECG-and-text AI models only compare a whole waveform to a whole report, which misses which part of the trace actually points to which diagnosis. The team also found that about 55% of standard clinical reports in the MIMIC-ECG dataset skip waveform descriptions entirely, so they used large language models to fill in that missing detail, then ran a separate check to catch and discard fabricated additions. To stop the system from wrongly penalizing reports that share a diagnosis but use different wording, they added a similarity matrix that corrects for those near-duplicates during training.
Zero-shot ECG interpretation, reading a heartbeat trace without task-specific training, has been a rough approximation until now, because global matching throws away the local evidence a cardiologist actually relies on. Tested across six datasets, FOCAL reportedly sets a new state-of-the-art in zero-shot prediction and linear probing, suggesting the fine-grained approach captures signal that coarser models discard.
Using an LLM to patch holes in medical records and then double-checking it for hallucinations is a tacit admission that the data underlying most ECG AI research is messier than the benchmark numbers let on.