Auscultation at the Edge: CNN-Powered Lung-Heart Sound Diagnosis for Low-Cost Telehealth Deployment

Abstract

Respiratory and cardiovascular diseases represent significant global health burdens, particularly in resource-constrained settings where access to specialized medical expertise is limited. This paper presents a comprehensive machine learning pipeline for auscultation-based diagnosis that converts short electronic stethoscope recordings into spectrogram- and chromogram-based tensors, trains a compact convolutional model with a fused scalar-audio head, and outputs multi-condition diagnostic labels. The system addresses asthma, COPD, pneumonia, heart failure, lung fibrosis, pleural effusion, and normal findings using a dataset comprising 112 participants recorded with a 3M Littmann electronic stethoscope across standardized chest zones. Preprocessing includes 2.5-second segmentation, 4 kHz resampling, denoising, and multi-channel feature construction to support robust learning from nonstationary respiratory and cardiac signals. The proposed architecture achieves strong performance on held-out data with overall accuracy, precision, recall, and F1-score all approximately 0.89, demonstrating particular resilience across minority classes. The complete system is operationalized on embedded hardware using a Raspberry Pi with a USB audio interface and a touch LCD for on-device recording, inference, and user feedback. This work illustrates a practical pathway from clinical audio machine learning to real-world telehealth deployment at the edge, offering potential for improved healthcare accessibility in underserved regions.