ECG arrhythmia classification via wavelet-driven feature extraction and swarm-optimised gradient boosting

Cardiovascular diseases have been the primary contributor to deaths worldwide, and hence, the need to detect arrhythmia from Electrocardiogram signals in a precise and efficient manner is a critical problem in the medical community. This work presents a lightweight and computationally efficient framework that integrates Discrete Wavelet Transform (DWT)-based statistical features, ECG morphological descriptors, and Artificial Bee Colony (ABC)-optimized eXtreme Gradient Boosting Machine (XGBM) classification for ECG beat analysis. This work has been implemented using the popular MIT-BIH Arrhythmia Database, which has 100,674 instances of ECG beats, divided into five AAMI classes, with a severe level of class imbalance, where 89.4% instances belong to the Normal class. ECG signals have been pre-processed using a seven-stage algorithm, including Butterworth high-pass filtering, notch filtering, Pan-Tompkins R-peak detection, beat segmentation, and normalisation. Then, a three-level Haar transform is implemented, and 32 statistical features have been extracted from the DWT decomposition, along with 32 morphological features, forming a 64-dimensional vector. The proposed ABC algorithm with 8 bees and 8 iterations optimizes the six XGBM model hyperparameters using a balanced fitness function of accuracy and macro F1-score and converges at the optimal fitness value of 0.8211. The proposed ABC-XGBM model has a classification accuracy of 95.14%, a macro F1-score of 0.948, a macro AUC of 0.983, Matthews Correlation Coefficient of 0.925, and G-Mean of 0.932 with class-wise AUC values > 0.94. An ablation study has shown that the proposed DWT adds +3.7% and the proposed ABC optimization adds +1.14% in accuracy improvement. Five-fold cross-validation has shown a stable performance with a mean accuracy of 0.952 ± 0.001 at a time complexity of 1.0 ms per sample without the dependency of the GPU. The proposed framework is better than the other deep learning models such as CardioAttentionNet with a classification accuracy of 91.20% and the proposed transformer-based classifier with a classification accuracy of 90.50%.