Ablation Driven CNN-BiLSTM-Attention Network With SHAP for ECG Arrhythmia Classification

The study examines both ECG classification performance and interpretability: how individual network components contribute to identifying heartbeat patterns, and which regions are associated with predictions.

This research explores ECG arrhythmia classification using a hybrid CNN-BiLSTM-Attention network. The study examines both classification performance and model interpretability, investigating how different network components contribute to identifying heartbeat patterns.

The architecture combines a convolutional neural network for extracting local ECG features, a bidirectional LSTM for capturing temporal patterns, and an additive attention mechanism for refining feature representations. The experiments use 109,446 annotated heartbeats from the MIT-BIH Arrhythmia Database, grouped according to AAMI EC57, with a test set of 21,892 beats.

The proposed model is compared with five baselines: CNN-only, BiLSTM-only, XGBoost, support vector machine, and k-nearest neighbors. On the study’s test set, the paper reports 96.84% accuracy, a macro F1-score of 0.8666, and a macro ROC-AUC of 0.9921.

Ablation experiments investigate the contribution of individual model components. SHAP explanations highlight ECG regions associated with predictions, including the QRS complex, helping make the model’s behaviour more understandable.

The paper was accepted at BECITHCON 2026 under Paper ID 530.

The architecture combines a CNN for local ECG features, a bidirectional LSTM for temporal patterns and additive attention. Ablation experiments examine the contribution of individual components, and the study compares the model with five baselines. SHAP explanations help inspect the regions associated with predictions.

The paper was accepted at BECITHCON 2026 under Paper ID 530. The reported test-set measurements and experimental context are documented in the research overview.