A deep learning method for ECG heartbeat classification based on Gaussian features

Electrocardiography (ECG) is a vital tool for diagnosing and screening cardiovascular diseases. Accurately extracting ECG signal features is key to improving the accuracy of arrhythmia classification. To address the limitations of traditional methods in terms of feature extraction precision and discrimination capability, this study proposes an ECG signal classification method based on Gaussian features. First, the R-peak is identified via the visibility graph algorithm, and its prominence is computed to locate the onset and offset of the P, Q, R, S, and T waves across the heartbeat waveform. Five Gaussian components are then fitted to the segmented signal, yielding 16 highly discriminative features for classification. Furthermore, an AMH–BiLSTM architecture is developed by integrating convolutional neural networks, bidirectional long short-term memory, and multi-head attention mechanisms, with two fully connected layers performing the final classification task. To address class imbalance in the MIT–BIH Arrhythmia Database, the SMOTE algorithm is employed to oversample minority classes. Moreover, over 10 independent random seeds, the proposed method achieves a mean classification accuracy of 97.93%, precision of 98.10%, recall of 97.92%, and F1-score of 97.99%, with a mean AUC of 99.90%. Compared to the baseline model under the same experimental setup, our proposed method has shown some improvements. In addition, we also provide reference comparisons with recent representative methods, and our method demonstrates competitive classification performance under comparable evaluation protocols.