Development of multihead cross-correlation attention fusion-based deep learning model for myocardial infarction detection using ECG signal and ESMOTE data balancing

Cardiovascular Diseases (CVDs) has become one of the leading causes of mortality in the world, with Myocardial Infarction (MI) being a serious cardiac disease that demands timely diagnosis to reduce fatality rates. ECG signals are largely used detecting MI because of their ability to capture the heart’s electrical activity and their non-invasive behavior. Nevertheless, the existing diagnostic models have some challenges such as data imbalance, noise artifacts and ineffective feature retrieval. These problems adversely impact the reliability and accuracy of the MI detection models. Nowadays, deep learning systems have proved enhanced potential in healthcare applications by allowing automatic feature learning and accurate classification. However, the effectiveness of these systems highly based on the quality of the signal pre-processing, effective feature extraction models and balanced datasets. To resolve these problems, a multi-head cross-correlation attention fusion-enabled deep learning model is proposed in this work for MI detection. At first, the essential ECG signals are collected from the benchmark database and given to a pre-processing stage to enhance the signal quality. Subsequently, the Enhanced Synthetic Minority Oversampling Technique (ESMOTE) model is adopted by improving the concept of the existing Synthetic Minority Oversampling Technique (SMOTE) technique for addressing the class-imbalance issue. Further, the balanced data is subjected to the feature extraction stage, where the significant features, such as wave features (F1), spectral features (F2) and deep features (F3) using the Spatial and Temporal Attention-based Sparse Autoencoder (STA-SA), are extracted. Finally, all these resultant features are fed into the Multihead Cross Correlation Attention Fusion-based Adaptive Feature-assisted Knowledge Distilled Temporal Convolution Network (MCCAF-AF-KDTCN), in which the three feature sets are fused by the Multihead Cross Correlation Attention Fusion (MCCAF), thus obtaining an adaptive feature. Then, the resultant adaptive fused feature is provided to the KDTCN model to achieve the classification outcome for MI disease. For further improvement, the hyperparameters of the KDTCN model are selected optimized with the help of Phase Transition-based Lyrebird Optimization (PTLO). The performance analysis is finally performed for the suggested model over traditional methods. The accuracy of the developed PTLO-MCCAF-AF-KDTCN is 95.78%, whereas the existing models, such as “Long Short-Term Memory (LSTM), Bidirectional Gated Recurrent Unit” (BiGRU), Convolutional Neural Network (CNN), and MCCAF-F-KDTCN, achieve 82.28%, 90.86%, 82.73% and 92.14% accuracy rates for the linear activation function. The results proved that the implemented MCCAF-AF-KDTCN is more effective for MI detection than the traditional approaches.