A CNN-based Local-Global Self-Attention via Averaged Window Embeddings for Hierarchical ECG Analysis

Electrocardiograms (ECGs) are multivariate time series where clinically relevant patterns span both local waveform morphology and long-range rhythm structure. We introduce LGA-ECG, a hierarchical Transformer architecture that integrates convolutional inductive biases into self-attention. Queries are extracted from overlapping local windows to retain morphological fidelity, while keys and values are globally derived to enable full temporal context. This design eliminates the need for explicit positional encodings by leveraging convolutional locality. On the CODE-TEST benchmark, LGA-ECG achieves a macro F1-score of 0.885, recall of 0.872, and precision of 0.907, outperforming CNN and Transformer baselines. Ablation studies confirm the effectiveness of combining local queries with global key-value pairs.