Lorenz-PSO Optimized Deep Neural Network for Enhanced Phonocardiogram Classification

Phonocardiogram is a crucial functional diagnostic tool in cardiology since cardiovascular disorders kill most people worldwide. Despite deep learning's popularity, dataset imbalance, signal feature repetition, and noise volatility hurt classification models. Two modifications of Lorenz chaotic system hybrid with PSO are given to improve optimal hyperparameter selection of deep learning architecture EfficientNet-To. A public dataset validates the Lorenz-PSO Optimized Deep Neural Network (LPODNN) framework for multiple disease classes of PCG signals. Compared to Mel Frequency Cepstral Coefficient and Discrete wavelet transforms feature extraction techniques hybrid with support vector machine, deep neural network, and centroid-based kth nearest neighbor classifiers, modification-I and modification-II achieve 99.25% and 99.5% accuracy, respectively. The results are also compared to a fusion of instantaneous frequency-based features with random forest & KNN classifiers and other state-of-the-art methods on accuracy, precision, recall, F1-score, and fitness value. The suggested approach is reliable since fitness value, loss function, and statistical measures are profiled on 100 independent runs. Different variants of Lorenz-PSO combined with deep learning can solve telemedicine and remote diagnosis problems.