Aims: Early and accurate ECG interpretation is critical, yet expert annotation remains costly and time-consuming to obtain. Existing ECG self-supervised learning (SSL) frameworks typically focus on either reconstruction or contrastive learning objectives in isolation, overlooking fine-grained intra-beat morphology and the anatomical organisation of the twelve leads. We propose two complementary extensions to a masked ECG language modelling backbone to address these limitations.
Methods: A syllable tokenizer decomposed each heartbeat into three sub-beat segments, capturing morphological detail at the level of individual waveform components e.g. T wave. A cross-plane contrastive learning objective exploited the anatomical grouping of ECG leads into frontal and transverse planes, encouraging alignment of complementary cardiac projections. Pre-training proceeded in two stages: vector-quantised heartbeat reconstruction followed by masked ECG sentence pre-training, with cross-plane contrastive learning applied simultaneously (Figure 1). Models were pre-trained on MIMIC-IV-ECG (N≈800,000) and evaluated on PTB-XL (N≈21,000) Form and Rhythm tasks via linear probing at 1%, 10% and 100% of available training labels.
Results: Both contributions yielded consistent improvements over the state-of-the-art ECG language modelling framework in the low-data regime. Syllable tokenization improved Rhythm AUC from 76% to 85% and Form AUC from 63% to 70% at 10% of training labels. Cross-plane contrastive learning improved Form and Rhythm AUC by 3.5 and 4.2 points respectively at 1% of training labels. The combined model, with both contributions, outperformed ST-MEM across all training data fractions.
Conclusion: Morphologically-informed tokenization and anatomically-grounded contrastive learning improve data efficiency in ECG representation learning, with particular relevance for clinical settings where expert annotation is scarce.