A multi-stage data preprocessing framework for ECG arrhythmia classification

Arrhythmia classification from ECG signals has been widely attempted by researchers using conventional feature-based methods and deep learning approaches. The UCI arrhythmia dataset is widely used for evaluating such classifiers. however, its missing values, outliers, and class imbalance present significant preprocessing challenges that can adversely affect classification performance and interpretability. The focus of the majority of the works is on achieving a fast, accurate, and reliable classifier; however, little effort has been made toward preprocessing the data. This study proposes a multi-stage, class-aware preprocessing framework for robust arrhythmia classification using the UCI dataset. The proposed framework systematically addresses data quality issues and evaluates the contribution of individual preprocessing modules across multiple classification algorithms, including low-complexity support vector machines. Experimental results demonstrate that each preprocessing stage contributes significantly to overall classifier performance. The proposed framework, yielding the best performance with an accuracy of 88.51%, with 0.8905 precision, 0.8851 recall, and 0.8834 F1-score. Furthermore, PCA achieved an accuracy of 85.14% with the added benefits of dimensionality reduction and noise mitigation. These results show the effectiveness of the proposed framework for arrhythmia classification, particularly when dealing with missing, noisy, and imbalanced ECG data.