Multi-Label Classification for Automated Diagnosis of Cardiac Abnormalities in ECG Readings

John Chauvin1, Soufiane CHAMI1, Rabie Fadil1, Sandeep Singhal1, Kouhyar Tavakolian2
1University of North Dakota, 2Associate Professor


Abstract

The electrocardiogram (ECG) is a non-invasive diagnostic tool that is used to diagnose a variety of cardiac abnormalities, many of which can lead to sever complications (including death) if not identified and treated as early as possible. Unfortunately, deciphering ECG recordings requires a high degree of skill and training and is most often done manually. Our aim is to automate the process of diagnosing cardiac abnormalities with the help of a multi-label classifier (i.e., one that can accurately assign one or more labels to the data). We use an ensemble classification approach leveraging both convolutional neural networks (CNNs) and traditional machine learning models to maximize classification accuracy. Features for the traditional models are derived from various aspects of segmented heartbeats and from heart rate variability metrics. A separate meta-learner is trained to combine the results of the individual classifiers in an optimal manner. As a starting point, we leveraged wavelet analysis to locate the QRS complexes in the Lead II readings and segment the ECG data into individual heartbeats. A multi-label K-Nearest Neighbors (MLKNN) model was then trained to assign one or more labels corresponding to either normal cariological conditions or to one of eight different cardiac ailments. Our team participated in PhysioNet Computing in Cardiology Challenge 2020 under the name BERCLAB UND. During training we evaluated the proposed model to yield an AUROC score of 0.500, an AUPRC score of 0.119, an accuracy of 0.799, an F_measure score of 0.099, an F_beta score of 0.103, and a G_beta score of 0.036 based on 5-fold cross validation. The Challenge submission system provided an F_beta score of 0.095 and a G_beta score of 0.030.