Stacking Deep Neural Networks to Detect Multiple Types of Cardiac Arrhythmias

Heart arrhythmias are associated with serious cardiovascular diseases and can result in fatal outcomes if not diagnosed early. Electrocardiograms (ECG) are generally used to diagnose heart arrhythmias. Prior studies employ deep learning architectures including Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and Transformer based methods to diagnose arrhythmias from ECG signals, achieving promising results. However, most of the deep learning models when used individually exhibit inherent limitations. CNNs can learn spatial features effectively but is limited in learning temporal characteristics. LSTMs capture temporal characteristics, but struggle with capturing long-range dependencies. Transformer models are effective at modeling long-range dependencies but they are often prone to overfitting. To address the limitations of individual models while leveraging their feature learning capabilities, this work proposes a two-level stacked ensemble framework for ECG arrhythmia classification. At the first level, deep learning models including a Deep Neural Network (DNN), CNN, LSTM, and Transformer models are trained as base learners while at the second level, the predictions generated by the base learners are combined and passed to the Multilayer Perceptron (MLP) which acts as the meta-learner. MLP learns how to weigh the predictions of the first level models to generate improved final predictions. The combination of these models in a stacked ensemble framework helps exploit the strengths of all models, enabling fusion of spatial, temporal, and long-range dependency features from the ECG signal. The proposed framework is evaluated on the MIT-BIH and INCART arrhythmias databases to classify the input ECG signals into five arrhythmia categories. Comparison against the base models and prior approaches demonstrates that the proposed stacked ensemble model outperforms the base models and state-of-the-art (SOTA) techniques, achieving an F-score of 99.79% on MIT-BIH and 99.62% on INCART dataset.