| Abstract [eng] |
According to the World Health Organization, cardiovascular diseases are the leading cause of death in the world. Artificial intelligence classification algorithms can help detect heart disease at an early stage. The goal of this work is to develop an algorithm for classifying heartbeat audio recordings (normal / abnormal signal) and compare the balanced accuracy of the resulting model with existing solutions. One of the main problems in phonocardiogram classification field is the limited amount of training data. This work uses the “2016 CinC/PhysioNet Challenge” dataset. The main problem in the dataset is data imbalance. There are almost 4 times more normal audio signals than abnormal ones. 3 different classification models (support vector classifier (SVM), random forests (RF) and a deep neural network based on convolutional neural networks (CNN)) are used to classify heart sounds and the results obtained are compared with existing solutions. Experiments are performed using 3 different input types: using segmentation, when audio recordings are segmented into medically recognizable heart sound segments (S1, S2, S3 and S4) and features are formed from the resulting segments, which are used as input to the models; without using segmentation, where inputs are obtained by calculating the averages of MFCC frames; using augmentation, where additional synthetic abnormal audio recordings are appended to the dataset to eliminate data imbalance and the averages of MFCC frames are used as input. When using augmentation and eliminating data imbalance, all models classified phonocardiograms 5-10% more accurately. The best result was achieved using a CNN-based deep neural network – 0.94 MAcc, quite similar result in terms of accuracy was also obtained using the SVM method – 0.9392 MAcc. Both CNN and SVM methods managed to achieve higher classification accuracy than the best existing solutions examined in the literature. |
