Beyond MIT-BIH: Failure Modes of Beat-Level ECG Arrhythmia Classification at Scale

Angus Nicolson1, Riccardo Lunelli1, Samuel Martin Proell1, Axel Bauer2, Clemens Dlaska1
1Digital Cardiology Lab, Medical University of Innsbruck, 2University Clinic of Internal Medicine III, Medical University of Innsbruck


Abstract

Beat-level ECG arrhythmia classification is benchmarked almost exclusively on small datasets: the majority of deep learning studies rely on MIT-BIH, a 48-recording, single-centre database. Whether these models generalise beyond this dataset is largely unknown. We evaluate this directly, fine-tuning xECG, an xLSTM-based foundation model, on INCART and MIT-BIH (inter-patient split, AAMI classes N, V, S, F, Q) and assessing beat-level predictions at scale on MIMIC-IV-ECG (706,073 recordings).

This exposes two distinct failure modes. First, patch-based models process fixed-length windows, so a single beat may span consecutive patches, causing double-counting. Suppressing consecutive same-class predictions treats contiguous same-class predictions as a single beat, but this heuristic fails at elevated heart rates when the RR interval approaches the patch duration. On MIMIC-IV-ECG, a patch-only model achieves beat-count Pearson r=0.678; under tachycardia (HR>120 bpm), it collapses to r=−0.420. Training R-peak detection jointly with beat classification resolves this: overall r rises to 0.961, tachycardia r to 0.776. On MIT-BIH, the joint model achieves a macro F1 of 0.599, V-class F1 of 0.973, and R-peak F1 of 0.996 (150 ms tolerance).

Second, and independently, distribution shift degrades precision substantially. On CPSC 2018 (6,871 12-lead recordings), converting beat-level predictions to recording-level ventricular ectopic beat detection yields sensitivity 0.930 but a PPV of 0.501. False positives are dominated by right bundle branch block, atrial fibrillation, and premature atrial contractions. On MIMIC-IV-ECG (F1=0.726, sensitivity=0.826, PPV=0.648), 52.5% of false positives contain conduction defects and 25.8% atrial fibrillation. Joint R-peak detection does not address this failure.

These results demonstrate that standard beat-level benchmarks substantially overstate clinical utility and that robust arrhythmia classification requires both reliable beat localisation and training data that reflects real-world rhythm diversity.