Comparing Top-K and Manual Feature Selection for PCG Signal Classification using XGBoost

Cardiovascular diseases (CVDs) are the leading cause of death, taking 18 million lives annually. Heart auscultation, the process in which clinicians record phonocardiogram signals to detect potential CVD abnormalities, plays an instrumental role in early detection and prevention of CVDs. However, traditional auscultation brings challenges such as requiring extensive clinician experience and skill. Therefore, this paper proposes an algorithm for Computer-Aided Auscultation, based on combining an ensemble gradient-boosting machine learning algorithm called XGBoost and various signal processing methods. 104 signals from clinics and in-home visits were acquired from the 2016 PhysioNet Challenge Dataset. After various preprocessing steps such as downsampling, normalization, and bandpass filtering, 25 timefrequency and spectral density features such as Spectral Power Density and S1 peak interval times were identified. Then, strong (highly predictive) features, especially those with greater than 0.7 correlation scores, were selected by utilizing boxplots and a correlation matrix. On the other hand, the ten best features were chosen using gradient-boosting sequential selectors. Finally, both sets of features were fed into multiple models like Logistic Regression, KNN, and LDA, and the anticipated XGBoost, which yielded a 92 % accuracy and a 93 % recall with the 10 manually selected features at a 90 % to 10 % training vs. test split. These promising results are pivotal in developing an accurate method to reduce clinical stress for heart auscultation, potentially saving lives.