Squeeze Excitation Network with MLP Classifier and Duck Swarm Optimization Based LU-Net for Heart Disease Prediction Using ECG Signal*

Heart disease can often be detected through analysis of ECG signals, which record the electrical activity of the heart. Current machine learning algorithms frequently face challenges such as overfitting, which arises from insufficient training data, as well as inconsistencies in ECG signal quality among diverse patient populations. ECG signals are first collected from the ECG heartbeat categorization dataset. Multivariate Dynamic Mode (MDM) decomposition technique, which dissects multivariate time-series data into spatial modes and temporal dynamics, effectively capturing both low and high-frequency elements of ECG signals. Following this, the LUNet method is employed to eliminate noise, while Duck Swarm Optimization (DSO) fine-tunes the batch size and learning rate for improved model performance. After denoising, the ECG signals are segmented to extract R-peak values using the Complex-Pan-Tompkins-Wavelets (CPTW) algorithm. Finally, the segmented features are classified using Squeeze Excitation Networks (SENet) in combination with a Multilayer Perceptron (MLP), facilitating the diagnosis of heart diseases by categorizing them as normal or abnormal. The proposed model outperforms current alternatives, achieving impressive metrics such as 97.24% accuracy, 95.78% PPV, 95.76% hit rate, 93.68% specificity, a miss rate of 0.23%, measuring 59 dB for PSNR and 46 dB for SNR, a MAE of 0.0036.