New Framework Links Heart Scans to Medical Reports with Fine-Grained Precision

Interpreting electrocardiograms (ECGs) requires connecting detailed waveform patterns to specific cardiovascular diagnoses—but this alignment problem has proven difficult for current AI systems. A new research framework called FOCAL tackles this challenge by establishing precise, localized matches between ECG segments and individual diagnostic tags found in patient reports.

The Core Problem#

Existing approaches to AI-assisted ECG interpretation rely on global representations, treating an entire heart scan as a single unit rather than examining the relationship between specific waveform sections and particular pathological findings. This broad-brush approach misses the nuanced connections that clinicians naturally observe. Complicating matters, nearly 55% of ECGs in standard medical datasets like MIMIC-ECG arrive with reports lacking explicit waveform descriptions, making it harder for systems to learn these fine-grained associations.

How FOCAL Works#

The framework uses Optimal Transport, a mathematical technique for measuring and aligning distributions, to link localized ECG patches directly to individual diagnostic tags. Because this tag-level matching can inadvertently treat different reports with overlapping diagnoses as misaligned, FOCAL introduces a semantic similarity matrix to guide the learning process and correct these errors. Additionally, to overcome the shortage of detailed waveform text in clinical records, the system applies large language models (LLMs) in a coarse-to-fine enrichment pipeline—generating missing semantic details while using a verification step to filter out AI hallucinations.

Validation and Impact#

Testing across six datasets shows that FOCAL achieves state-of-the-art results in both zero-shot ECG prediction (interpreting new data without prior labeled examples) and linear probing (using learned features with minimal retraining). The research demonstrates that precise alignment between waveform regions and clinical findings improves the model's ability to generalize to unseen cases and extract interpretable diagnostic signals from raw electrocardiographic data.