Age-Based Heart Sound Segmentation for Accessible Pediatric Cardiovascular Disease Screening

Digital stethoscopes provide a possible cost-effective solution to accessible screening of cardiovascular diseases in low-to-middle-income countries. Heart sound segmentation is an essential step in computer-aided screening. This paper examines the underlying adult-based assumptions and presumptions of state-of-the-art heart sound segmentation algorithms, and then proposes an age-based heart sound segmentation to provide higher accuracy performance for pediatric phonocardiograms. CirCor DigiScope Dataset was utilised, containing 3163 heart sound recordings from 942 pediatric patients ranging from neonate to young adult age groups. Compared to existing adult-based assumptions and presumptions, 5.4%-80% of patients were outside the expected heart rate range, and expected S1 and S2 duration distributions showed an overlap of 53.5%-96.3% and 76.3%-91.3% respectively, with younger age groups showing the largest differences in most cases. Additionally, the assumption of the linear relationship between mean systole duration with total systolic interval was weaker for younger age groups. Utilising a pediatric-based denoising algorithm and appropriate modification of major parameters and relationships within an existing heart sound segmentation algorithm, it was shown that S1 and S2 heart segmentation F1-Scores improved in all age groups from 69.0%-93.7% to 91.4%-99.8%.Clinical relevance- Accurate heart sound segmentation is a necessary preliminary step for automated clinical decision assistance tools for cardiovascular disease screening.