• Rutuja Jadhav
• Peeyush Pareek

Automated electrocardiogram (ECG) arrhythmia classification is essential for scalable, non-invasive cardiac monitoring, yet many reported models are evaluated under intra-patient protocols that overstate real-world performance. This paper presents an inter-patient, AAMI EC57-compliant five-class arrhythmia classification pipeline evaluated on the MIT-BIH Arrhythmia Database under the De Chazal DS1/DS2 split. ECG beats are denoised via bandpass filtering and wavelet thresholding, segmented into fixed windows around annotated R-peaks, and represented using a 67-dimensional handcrafted feature set reduced to 40 features via Minimum Redundancy Maximum Relevance (MRMR) selection, alongside raw waveform input for deep models. Three deep learning architectures, a 1-D CNN, a BiLSTM, and a proposed Multi-Scale CNN-BiLSTM with dual attention, are trained on a two-phase SMOTE-ENN-balanced training set and combined via soft-voting, weighted-voting, and stacking ensembles. The stacking ensemble achieves the best inter-patient test performance, with 94.45% accuracy and a macro-F1 of 76.66% across the five AAMI classes, and Optuna-based hyperparameter tuning further improves the standalone proposed architecture's macro-F1 from 62.99% to 67.12%. External validation on the independent INCART database, performed without retraining, shows a substantial generalization gap (macro-F1 of 35.79%), highlighting cross-database robustness as an open challenge. The supraventricular class remains the dominant source of error across all models and both databases. Results are reported transparently, including configurations where ensembling did not improve on the best individual model, to support reproducible benchmarking in inter-patient ECG arrhythmia classification.

ECG arrhythmia classification, inter-patient paradigm, AAMI EC57, MIT-BIH, INCART, CNN-BiLSTM, dual attention, ensemble learning, SMOTE-ENN, MRMR feature selection

"An Efficient Multi-Scale CNN-BiLSTM with Dual Attention Mechanism for Inter-Patient ECG Arrhythmia Classification: Evaluation on MIT-BIH and INCART Databases", International Journal of Emerging Technologies and Innovative Research (www.jetir.org), ISSN:2349-5162, Vol.13, Issue 7, page no.b148-b171, July-2026, Available :http://www.jetir.org/papers/JETIR2607122.pdf

"An Efficient Multi-Scale CNN-BiLSTM with Dual Attention Mechanism for Inter-Patient ECG Arrhythmia Classification: Evaluation on MIT-BIH and INCART Databases", International Journal of Emerging Technologies and Innovative Research (www.jetir.org | UGC and issn Approved), ISSN:2349-5162, Vol.13, Issue 7, page no. ppb148-b171, July-2026, Available at : http://www.jetir.org/papers/JETIR2607122.pdf