Optimized Hybrid Deep Network for Heartbeat Sound Classification Using Dolphin Guided Modified Blackman Windowing Segmentation

Early detection of cardiovascular diseases (CVD) is critical for improving patient outcomes and preventing severe cardiac complications. However, traditional heart sound classification methods often suffer from limitations such as low accuracy, inefficient signal processing, and high computational cost. These shortcomings hinder the timely and accurate diagnosis of CVD. To address these challenges, a novel two-phase evaluation framework is proposed. In the first phase, a strong method for breaking down and identifying features is created by combining Dolphin Echolocation Optimization (DEO) with the Onset technique, which is improved using Modified Blackman Windowing (MBW) to reduce unwanted noise in the data. Mel Frequency Cepstral Coefficients (MFCC), DEO, and onset features are jointly utilized to accurately segment heart sounds for each cardiac cycle. In the second phase, a hybrid deep learning architecture combining Separable Convolutional Neural Network and Bidirectional Long Short-Term Memory (SEPCNN-BiLSTM) is proposed to improve classification accuracy and minimize prediction loss. The proposed model demonstrates superior performance over existing approaches, achieving a peak classification accuracy of 94.8% at batch size 32, with a total model training time of 90 s and an inference response time of 140 ms per sample, outperforming all evaluated baseline architectures.