Abstract:
Heart sounds play a vital role in assessing cardiac health. This article introduces a novel congenital heart disease classification algorithm based on the fusion of features utilizing variable logic and linear predictive cepstral coefficients. This approach aids in the extraction of profound pathological characteristics from heart sounds. The algorithm begins by denoising and extracting envelopes from the heart sounds. It then proceeds to perform variable logic operations, labels, and transforms the data into analyzable measurement data. Furthermore, it calculates the linear predictive cepstral coefficients of the signal for feature fusion. Finally, a binary classification of congenital heart disease is conducted using machine learning classifiers, specifically random forest (RF), XGBOOST and LIGHTGBM. The study utilized a total of 4 000 heart sound samples, resulting in an average accuracy of 0.913 8 in distinguishing between normal and abnormal heart sounds. Importantly, this algorithm eliminates the need for segmenting heart sounds by cardiac cycles, significantly streamlining the analysis process, and holds promise for congenital heart disease screening.