An Innovative Framework for Heart Sound Classification Integrating Adaptive Fuzzy Rank-Based Ensemble of Transfer Learning Models with Multi-Dimensional Features

Abstract

Heart sound identification for diagnosing cardiovascular diseases is a complex challenge due to the intricate, spectro-temporal characteristics of conditions like Aortic Stenosis, Mitral Regurgitation, and multi-valvular disorders. Conventional methods frequently inadequately encompass the complete range of diagnostic characteristics, depending on fragmented or simplistic approaches that lack generalizability across varied patient demographics, clinical settings, and recording circumstances. To address these limitations, we offer an innovative, multi-dimensional feature fusion framework that integrates Wavelet Scattering Transform (WST) and Mel-Frequency Cepstral Coefficients (MFCC) to capture both temporal stability and perceptually optimal frequency patterns from cardiac sounds. This method is further refined by an adaptive fuzzy rank-based ensemble technique utilizing the Gompertz function, which dynamically modifies model weights according to confidence metrics, hence assuring more dependable and precise predictions amidst fluctuating clinical uncertainty. We meticulously assess our model utilizing eight advanced fine-tuned transfer learning architectures across four feature extraction techniques (WST, MFCC, STFT, Multi-Dimensional) on the clinically validated BUET Multi-disease Heart Sound (BMD-HS) dataset. This dataset comprises 864 phonocardiogram recordings from 108 participants with echocardiographically confirmed diagnoses. The multi-dimensional feature fusion method attains 97% accuracy with EfficientNetB2, whereas the adaptive fuzzy ensemble strategy achieves 98% accuracy, surpassing both individual models and conventional ensemble methods. Moreover, Explainable AI with Audio-LIME offers transparent, clinically interpretable insights, achieving fidelity scores surpassing 0.85 and clinical relevance ratings exceeding 90%, facilitating the identification of critical time-frequency regions that hold diagnostic significance. This system establishes a novel benchmark for heart sound classification, enhances the differentiation of intricate valvular disorders, and provides dependable, evidence-based decision support for cardiovascular diagnosis in clinical settings. Our research illustrates the capability for resilient, scalable, and interpretable heart sound classification systems that can improve clinical decision-making and promote the integration of automated diagnostic tools in healthcare.