Fusing handcrafted and deep features for multi-class cardiac diagnostic decision support model based on heart sound signals

Among all causes of death, cardiovascular disease (CVD) is the leading one. In the medical field, the phonocardiogram (PCG) signal is widely used due to its cost-effectiveness and simplicity. We propose a hybrid classification approach for automatically classifying different cardiac auscultation classes using a PCG signal using handcrafted and deep learned features. Firstly, the PCG signal is split into equal segments using the windowing procedure. The power spectrum of a signal is constructed using l -spectrograms and then the novel modified dilated ResNet structure generates the deep features from the spectrogram images. Simultaneously, handcrafted frequency domain features are extracted from correspond windowed signals. After features are extracted, radial basis function (RBF) kernel in support vector machines (SVM) is used to classify various patterns. Moreover, the optimized SVM is performed using a modified drone squadron optimization (mDSO) algorithm to prevent overfitting. The proposed approach is 94.16% accurate for Michigan Heart Sound and Murmur Database (MHSDB) and 99.38% accurate for PhysioNet/CinC 2016 Challenge. The introduced hybrid model is more computationally efficient and accurate than earlier state-of-the-art strategies, based on the evaluation measures. Using the designed model, underserved and diverse communities can implement automated cardiac screening systems.