PRADA: Phase-Rhythm Adaptation of Time Series Foundation Models for ECG

Time series foundation models (TSFMs) have demonstrated strong transferability across diverse domains and can achieve competitive performance on electrocardiogram (ECG) diagnosis even with simple linear probing. However, their generic temporal modeling is not inherently aligned with the physiological structure of ECG signals: uniform patching segments cardiac cycles independently of cardiac phase, fragmenting characteristic P–QRS–T morphology across patch boundaries, while patch-based representations provide no explicit mechanism for preserving inter-beat rhythm information. We propose PRADA (Phase-Rhythm Adaptation), a parameter-efficient framework that adapts pretrained TSFMs to ECG by preserving both beat morphology and rhythm. PRADA first constructs a phase-locked representation by aligning individual beats around R-peaks. It then applies shared-bottleneck LoRA, using a shared low-rank bottleneck and scale-specific projections to jointly adapt original and beat-aligned ECG representations with a frozen backbone. Finally, a rhythm-conditional gate combines their class-wise predictions according to R-peak-derived rhythm statistics. Across four ECG datasets and seven diagnostic tasks, PRADA outperforms linear probing in all 28 backbone–task combinations across MOMENT, LPTM, Mantis, and SymTime, with absolute AUROC gains of up to 21.59 percentage points while training only 1.69\% of the backbone parameters. The largest improvements occur on morphology- and rhythm-dependent tasks, demonstrating the effectiveness of explicitly incorporating ECG phase and rhythm into TSFM adaptation.