Overcoming Class Overlap and Imbalance in ECG Detection and Classification: A Deep Attention-Based Model on MIT-BIH

ECG signals are essential for monitoring and diagnosing cardiovascular conditions, particularly arrhythmias, which can lead to severe complications if undetected. This study introduces an AI-based approach for arrhythmia classification using the MIT-BIH arrhythmia dataset, addressing challenges like class imbalance, class overlap, and intra-patient bias. To enhance data quality, the dataset was balanced using the Synthetic Minority Oversampling Technique (SMOTE) and augmented with Gaussian noise for minority class samples. A Conv1D-Attention network was employed during preprocessing to extract local ECG features and focus on key waveforms. Among the evaluated classifiers, decision tree, random forest, and support vector machine (SVM), the random forest achieved the highest accuracy of 91%. Although preprocessing reduced class imbalance and variance, a drop in performance was observed. This reflects a realistic evaluation scenario by preventing data leakage from similar ECG segments of the same patient in both training and test sets. Enforcing patient-independent segmentation compelled the model to generalize beyond individual patterns, a critical step for real-world applications. This study highlights the importance of rigorous evaluation protocols in biomedical machine learning. Combining data augmentation with attention-based feature extraction significantly enhances model generalizability, particularly in handling overlapping and imbalanced classes. This approach shows promise for developing reliable, patient-independent diagnostic tools for early arrhythmia detection in clinical settings.