Audio recordings of heart sounds, so-called phonocardiograms (PCGs), provide a useful basis for computer-aided diagnosis for early detection of congenital or rheumatic heart disease. Pre-screening of PCGs using machine learning algorithms could be useful where appropriate medical care is not available due to economic or social reasons. Such algorithms could also provide continuous bedside monitoring.
In the George B. Moody PhysioNet Challenge 2022, three classes are to be identified from PCG data, namely presence or absence of heart murmurs or 'unknown' for inconclusive diagnoses. We (team "listNto_urHeart") propose an algorithm including signal preprocessing prior to classification by a Long Short-Term Memory (LSTM) network. Preprocessing includes bandpass filtering, outlier removal, and percentile-based normalization. We employed an adaptive weighting scheme for training the LSTM to overcome the class imbalance. To focus on high-quality data, we used only the most audible location for the 'present' class. For the other classes, we chose uniformly at random a single location to not re-introduce class imbalance.
After training, the LSTM network will predict the 'present' or 'unknown' class, if any recording for the corresponding sample scores at least 75% of the probability mass in said class, respectively. This trade-off prefers referring healthy cases to a doctor over missing present heart disease. Our algorithm achieved a score of 1560 in the unofficial phase. Based on previous entries, we assume that the preprocessing is the crucial step for precise classifications due to low signal-to-noise ratio in many provided PCGs caused by talking, crying infants or other loud noises.
Our approach highlights the importance of appropriate preprocessing of PCG data and lays the ground work for automated PCG analysis.