HeartSound AI: A Deep Learning Platform for Automated Phonocardiogram Analysis Enabling Smartphone-Based Cardiac Screening System Architecture, Pilot Validation, and Pathway to Clinical Trials
Abstract: Background: Cardiovascular diseases remain the leading cause of global mortality (17.9 million deaths annually). Early detection through cardiac auscultation offers critical intervention opportunity, yet diagnostic accuracy varies widely (sensitivity 32-44% for clinically significant valvular heart disease). AI-powered phonocardiogram analysis presents a promising approach to democratize cardiac screening. Objective: To develop and conduct preliminary validation of HeartSound AI, a comprehensive deep learning platform for automated PCG analysis designed for smartphone deployment. Methods: We designed a modular pipeline integrating: (1) adaptive preprocessing with bandpass filtering (20-800 Hz); (2) signal quality assessment; (3) heart sound segmentation using Hidden Semi-Markov Models; (4) spectro-temporal feature extraction (MFCCs, CWT, Mel-spectrograms); (5) hybrid CNN-RNN classifier with attention; and (6) calibrated probability fusion. Development followed Good Machine Learning Practices guidelines. Evaluation utilized CirCor DigiScope dataset (5,272 recordings) and pilot clinical recordings (n=15). Results: On development data, the platform achieved weighted accuracy of 74.2% (95% CI: 71.8-76.6%) with murmur-present sensitivity of 81.3%. Heart sound segmentation demonstrated S1/S2 detection rates exceeding 94%. Real-time analysis was achieved in
Keywords: Phonocardiogram; Deep Learning; Cardiac Auscultation; Heart Murmur Detection; Software as Medical Device; Mobile Health; Point-of-Care Diagnostics; Valvular Heart Disease
How to Cite?: Sanjeeva Reddy Bora, "HeartSound AI: A Deep Learning Platform for Automated Phonocardiogram Analysis Enabling Smartphone-Based Cardiac Screening System Architecture, Pilot Validation, and Pathway to Clinical Trials", Volume 14 Issue 12, December 2025, International Journal of Science and Research (IJSR), Pages: 1949-1955, https://www.ijsr.net/getabstract.php?paperid=SR251223184722, DOI: https://dx.doi.org/10.21275/SR251223184722