Beat-Synchronous Tokenization for ECG Transformers

Summary

University of Minnesota group asks a narrow, well-controlled question: does tokenizing an ECG waveform at cardiac-cycle (beat) boundaries instead of fixed time patches change Transformer performance, holding everything else (encoder architecture, SSL objective, evaluation protocol) constant? They compare fixed patches against three beat-synchronous tokenizer designs — resampled beats, adaptive-pooled beats, and resampled beats plus explicit R-R interval timing — on two settings: 12-lead diagnostic classification (PTB-XL after MIMIC-IV-ECG masked pretraining) and single-lead rhythm classification (Icentia11k after patient-level contrastive pretraining). No wrist signal, no PPG, no accelerometer — this is a pure ECG tokenization ablation.