Beyond Black Boxes: Multi-Domain Soft Set Framework for Explainable ECG Arrhythmia Detection
Abstract
Electrocardiogram (ECG) arrhythmia classification remains critical for early heart disease detection, but current deep learning approaches sacrifice interpretability for performance gains. This limitation poses obstacles to clinical application. To address this challenge, a multi-domain soft set framework was introduced and achieved 96.9\% accuracy using only 35 interpretable features on the MIT-BIH Arrhythmia Database. Four machine learning classifiers (Random Forest, XGBoost, LightGBM, SVM) were systematically compared via 5-fold cross-validation, with LightGBM selected to provide the optimal balance between accuracy (96.9\%), computational efficiency (approximately 9 seconds per fold), and stability. To assess generalizability, cross-database validation was further conducted on the St. Petersburg INCART 12-lead Arrhythmia Database using the identical framework and parameter configuration, with all four classifiers achieving accuracies exceeding 96\%.Through extensive ablation studies, soft set features were shown to contribute an absolute improvement of 7.7\% over baseline statistical features (89.2\% vs. 96.9\%). SHAP explainability analysis confirmed the physiological interpretability of the framework, revealing that temporal dynamics (TDSS-thresh: 60.6\% significance) and frequency features (FTSS-std: 52.9\%) were the most discriminative features. The proposed method uses only 35 interpretable features compared to thousands of parameters in deep learning models. With 96.9\% accuracy, the framework offers competitive performance while maintaining full transparency.
Keywords
Arrhythmia classification, ECG, Interpretable machine learning, MIT-BIH database, Soft set framework, Soft set theory
References
- G. B. Moody, R. G. Mark, The impact of the MIT-BIH arrhythmia database, IEEE Eng. Med. Biol. Mag., 20(3) (2001), 45–50. https://doi.org/10.1109/51.932724
- S. Kiranyaz, T. Ince, M. Gabbouj, Real-time patient-specific ECG classification by 1-D convolutional neural networks, IEEE Trans. Biomed. Eng., 63(3) (2016), 664–675. https://doi.org/10.1109/TBME.2015.2468589
- Ö. Yildirim, A novel wavelet sequence based on deep bidirectional LSTM network model for ECG signal classification, Comput. Biol. Med., 96 (2018), 189–202. https://doi.org/10.1016/j.compbiomed.2018.03.016
- R. He, Y. Liu, K. Wang, et al., Automatic cardiac arrhythmia classification using combination of deep residual network and bidirectional LSTM, IEEE Access, 7 (2019), 102119–102135. https://doi.org/10.1109/ACCESS.2019.2931500
- R. Hu, J. Chen, L. Zhou, A transformer-based deep neural network for arrhythmia detection using continuous ECG signals, Comput. Biol. Med., 144 (2022), Article ID 105325. https://doi.org/10.1016/j.compbiomed.2022.105325
- N. Anjum, K. A. Sathi, M. A. Hossain, et al., A temporal transformer-based fusion framework for morphological arrhythmia classification, Computers, 12(3) (2023), Article ID 68. https://doi.org/10.3390/computers12030068
- Y. Xia, Y. Xu, P. Chen, et al., Generative adversarial network with transformer generator for boosting ECG classification, Biomed. Signal Process. Control, 80(1) (2023), Article ID 104276. https://doi.org/10.1016/j.bspc.2022.104276
- Y. Qiu,W. Chen, L. Yue, et al., STCT: Spatial-temporal conv-transformer network for cardiac arrhythmias recognition, International Conference on Advanced Data Mining and Applications, Springer, Cham, (2022), 86–100. https://doi.org/10.1007/978-3-030-95405-5_7
- H. Li, J. Han, H. Zhang, et al., Clinical knowledge-based ECG abnormalities detection using dual-view CNN-Transformer and external attention mechanism, Comput. Biol. Med., 178 (2024), Article ID 108751. https://doi.org/10.1016/j.compbiomed.2024.108751
- M. R. Islam, M. Qaraqe, K. Qaraqe, et al., CAT-Net: Convolution, attention, and transformer based network for single-lead ECG arrhythmia classification, Biomed. Signal Process. Control, 93 (2024), Article ID 106211. https://doi.org/10.1016/j.bspc.2024.106211