ECG-InterpBench: Benchmarking Interpretability of ECG Foundation Models with Matched-Scale Sparse Autoencoders

Yixuan Duan, Wei Qiu

Summary

ECG-InterpBench applies mechanistic interpretability tooling (matched-scale Sparse Autoencoders) to ECG foundation models. SAEs learn sparse overcomplete dictionaries from FM activations, producing monosemantic features that ideally map to human-recognizable ECG patterns (P-wave, QRS duration, ST elevation, T-wave inversion). "Matched-scale" ensures fair cross-model comparison.