Congenital heart diseases, a term that includes a variety of birth defects impacting the normal heart function, affects almost 1 in 100 babies. The presence of heart murmurs in phonocardiogram (PCG) signals is an indicator of these conditions. Successful identification of heart murmurs would reduce the chance of misdiagnosis and eventually help to design future treatments.
Here, we propose a new signature-based model to identify heart murmurs. Our approach treats the PCG signal as a path in time from which the signature method extracts useful features. These features, together with other available information, are fed into a gradient boosting algorithm, which is trained to classify the presence of murmurs for each recording location. The overall classification for each patient is optimised using the provided cost function.
For our initial benchmark submission we received a score of 2129, which corresponds to an average 10-fold cross-validation score of 2067 (sd=279). The subsequent model that included basic configuration of path signatures yields an average cross-validation test score of 2000 (sd=272).
Our early results show that the proposed application of path signature method, designed to extracts time-varying information from the sequential data, may lead to a significant model improvement, which we will further explore in the official phase.