Using Mel-Frequency Cepstrum and Amplitude-Time Heart Variability as XGBoost Handcrafted Features for Heart Disease Detection

Abstract

We have developed the XGBoost model to identify 27 heart pathologies within the challenge Will Two Do? Varying Dimensions in Electrocardiography: The PhysioNet/ Computing in Cardiology Challenge 2021. The technical part included several stages. At the first stage, the ECG was cut off to 10 seconds. At the second stage, resampling to frequencies 125 and 500 Hz was carried out and filtering in the 0.5-45 Hz bands. At the third stage, the features of HRV and symbolic dynamics were extracted from the signal with a sampling rate of 125 Hz. The melspectrograms were calculated based on a signal with a sampling frequency of 500 Hz. Then the features calculated for each lead were concatenated to obtain the final vector of features. We were faced with the task of constructing 27 independent binary classifiers, each of which defines a certain pathology. The fourth important step was to build balanced datasets for the algorithm. For the robustness of the models, the control groups for each contained almost all pathologies presented in the databases, except target disease. Our team Sunset scored 0.22, 0.21, 0.22, 0.21, 0.20 for the 12-lead, 6-lead, 4-lead, 3-lead, and 2-lead models, respectively, ranking 32 out of 39 teams for the first four lead combinations and 31 out of 39 teams for the last.