Autoregressive Whitening Filter for Detection of Coronary Artery Disease Based on Phonocardiography

Bjarke Skogstad Larsen1, Simon Winther2, Louise Nissen2, Axel Diederichsen3, Morten Bøttcher2, Johannes Struijk1, Mads Græsbøll Christensen1, Samuel Emil Schmidt1
1Aalborg University, 2Department of Cardiology Hospital Unit West, 3Department of Cardiology, Odense University Hospital


Background: Narrowing of the coronary arteries, which defines Coronary Artery Disease (CAD), can be discovered through analysis of phonocardiography (PCG) recordings. Previous work has found an average increase of power in frequencies below 200 Hz, and has shown that this can be used to distinguish between healthy (Non-CAD) subjects and subjects with CAD. However, spectral roll off of the PCG is steep (~40 dB/dec), meaning that spectral leakage might overshadow the very weak CAD related signal. Methods: Heart sound recordings from 1356 subjects, 258 CAD and 1098 Non-CAD, were pooled from three different studies. All recordings were acquired using the Acarix CADscor System at the fourth inter-costal space (IC4). The mid-diastolic segments were extracted using an automated algorithm, and an AR model was built from each segment. An average AR model was constructed as an average of the AR coefficients from the Non-CAD subjects. The inverse coefficients of the average AR model were then used as a whitening filter before power spectral analysis of diastolic segments. Results: The average diastolic power spectrum of Non-CAD subjects was white within 10.6 dB interval in the 10 to 800 Hz frequency range. The average diastolic power spectrum of CAD subjects was at least 5 dB higher than that of Non-CAD subjects in the 40 to 150 Hz frequency range, with a maximum of 7.7 dB at 68 Hz. Conclusion: Whitening of diastolic PCG segments emphasized the difference between Non-CAD and CAD subjects.