Hybrid machine learning models for enhanced arrhythmia detection from ECG signals using autoencoder and convolution features

Automated arrhythmia detection from electrocardiogram (ECG) signals is crucial and important for the early treatment of cardiac disease (CD). In this investigation, eight machine-learning models have been developed to identify improved ECG arrhythmia detection using two standard datasets (MIT-BIH Arrhythmia and the ECG 5000). In the first phase, two types of feature extraction schemes (autoencoder) and (Convolution) are used to obtain relevant features from ECG samples and subsequently, eight ML models are successfully trained and tested to find various performance matrices through simulation-based experiments. Then, the TOPSIS and mRMR ranking schemes are used to rank the ML models and identify the three best-performing models recommended for real-time arrhythmia detection. In this study, it is observed that for the same number of input features, models based on autoencoder features offer enhanced performance compared to those based on convolutional features. It is generally observed that the top identified hybrid model, Autoencoder Features with Neural Network (AEFNN) on the MIT-BIH dataset, achieves an accuracy of 97.96% and on the ECG5000 dataset, the hybrid model achieves an accuracy of 99.20%. This proposed model can be utilized for the early detection of arrhythmia, particularly in large-scale healthcare screening programs, thereby aiding in timely diagnosis and intervention. In this study, two types of features are used to model development in future work. Other relevant important features can be extracted from ECG samples, and those features can be used to develop accurate models to identify Heart disease.