Detection of Heart Murmurs in Phonocardiograms with Parallel Hidden Semi-Markov Models

We describe a recurrent neural network and hidden semi Markov model (HSMM) approach to detect heart murmurs in phonocardiogram recordings. This model forms the winning 'CUED_Acoustics' entry to the 2022 George B. Moody PhysioNet challenge. Segmentation of the phonocardiogram is a key pre-processing step for many heart sound algorithms. However, most previous work assumes that heart sound recordings only contain S1 and S2 sounds, leading to poorer segmentations of signals that contain a strong murmur. Our approach applies multiple HSMMs, each making different assumptions about a possible murmur, to produce multiple segmentations of the signal. We then compare the confidence of each HSMM's output to produce both a murmur classification and robust segmentation. Evaluated on the hidden test set, our algorithm achieved a weighted accuracy score of 0.776 on the murmur detection task (ranked 2nd of 40 teams, and just 0.004 below the top score). On the clinical outcomes task, the algorithm achieved a challenge cost score of 11144 (ranked 1st of 40 teams). The high performance on both tasks suggests the algorithm is sensitive to clinically significant murmurs. Compared to end-to-end models, the algorithm also provides interpretable results about their location and timing. This makes it a promising tool for symptomatic screening.