Coronary Artery Disease Detection from PCG signals using Time Domain based Automutual Information and Spectral Features

This paper proposes a quick, compact and cost-effective point-of-care stethoscope-based device that detects Coronary Artery Disease (CAD) from phonocardiogram (PCG) signals, i.e. Recordings of heart sounds, compared to existing methods which are either expensive or are unable to diagnose until the conditions too severe. PCG signals are extracted from patients using a condenser microphone mounted on a stethoscope and is followed by amplification and filtering. The signals are passed through the laptop using an audio jack and digitized. Thereafter they are segmented into the 4 states S1, systole, S2 and diastole using a Hidden Semi Markov Model (HSMM). Afterwards, the diastolic phases are isolated and both time and frequency domain features are analyzed. In the time domain, features are extracted using a nonlinear function, the Automutual Information. In the frequency domain, both high and low-frequency domain features were extracted. A Support Vector Classifier using a Radial Basis Function was trained on 190 recordings from the 2016 PhysioNet/Cinc challenge and obtained an accuracy of 0.74, indicating the combined use of both time and frequency measures from PCG signals could be viable. Such a product could be of great use to clinicians as a quick, inexpensive and primary means of checking whether or not a patient has CAD.