TriSpectraKAN: a novel approach for COPD detection via lung sound analysis

Abstract

Abstract This study aims to create an automated, accessible, and cost-effective diagnostic tool for chronic obstructive pulmonary disease (COPD). Traditional diagnostic methods are expensive, time-consuming, and require specialized equipment. The proposed TriSpectraKAN model leverages audio-based lung sound features to improve early diagnosis. TriSpectraKAN is a hybrid model combining spectral features and the Kolmogorov–Arnold Network (KAN) to analyze lung sounds using Mel-frequency cepstral coefficients (MFCCs), chromagram, and Mel spectrograms. Each sub-model focuses on a different audio feature, capturing unique sonic signatures. These features are merged through a hybrid network for comprehensive analysis. The model, trained on a COPD dataset, was deployed on a Raspberry Pi for real-time use. TriSpectraKAN achieved 93% accuracy, an F1 score of 0.98, precision of 0.97, and recall of 0.98. This multimodal approach captured a broad range of lung sound features, improving diagnosis accuracy compared to traditional methods. The integration of multiple audio features in TriSpectraKAN enhances COPD diagnosis, demonstrating the potential of AI and machine learning to transform respiratory disease diagnosis through accessible tools.