PCG Murmur Detection by GA ensemble of Scalogram Based Convolutional Neural Network and Random Forest Classifier

Muhammad Zoraiz Ramay and Muhammad Usman Akram
Department of Computer and Software Engineering, National University of Sciences and Technology, Islamabad 44000, Pakistan


The objective of the 2022 PhysioNet/CinC Challenge is to develop a classification algorithm based on phonocardiograms (PCGs) to predict whether heart murmur is present, absent or needs expert examination to decide. Performance evaluation function takes into account of false predictions and their associated costs. The good classifier not only reduces cost, time of detecting murmurs but will save lives by starting early treatment. Here in this paper, classification algorithm is developed and evaluated on test data.

In the pre-processing steps, a Butterworth bandpass filter is be applied to cut off high and low frequencies to reduce noise. Our algorithm makes 2D scalogram images generated on the application of the continuous wavelet transform on PCGs and then implements convolutional neural network (with several convolutional layers with padding, ReLu and several pooling layers). We then use random forest classifier with age, gender, i.e., demographic information and ensemble these two models using weightages. Weightages will be determined by genetic algorithm (GA).

Team name: MZR_CEME Physionet Challenge Unofficial Score: 2011