ECG Multi-Class Diagnostics Using a Hybrid CNN-Transformer with Self-Supervised Pre-Training

Automated ECG interpretation is held back by three practical problems. These are skewed class distributions in public corpora, a shortage of expert-labeled signals, and the need to model both local waveform morphology and beat-to-beat dependencies over a 10 -second window. We address these issues with a hybrid CNN-Transformer trained in two stages. First, the encoder is pre-trained on unlabeled 12-lead recordings using a combination of masked segment prediction (MSP) and NT-Xent contrastive learning, after which the classification head is attached and fine-tuned on the five PTB-XL super-classes. Pretraining lifts macro-F1 from 0.5845 to 0.6766 and macro-AUC from 0.8851 to 0.9293 relative to a supervised baseline trained under identical conditions. An ablation over the two SSL objectives shows that the contrastive term alone barely moves the baseline, whereas MSP contributes almost the entire AUC gain, and the two objectives combined are still noticeably stronger than either one in isolation.