Multi-class Classification of 12-lead Electrocardiogram Recordings to Assist in the Diagnosis of Cardiac Abnormalities

Jacob Kimball and Omer T. Inan
Georgia Institute of Technology


Abstract

Introduction:

The electrocardiogram (ECG) allows for the noninvasive assessment of many aspects of heart health. However, parsing ECG recordings is time-intensive. Automatic detection and classification of recordings by machine learning models can potentially reduce the time to diagnosis for each patient and lower the burden on healthcare providers.

Methods:

Using data provided by the PhysioNet/Computing in Cardiology Challenge 2020, clinically relevant features such as timing intervals and amplitude ratios were calculated for each of the 12 ECG leads per recording. Engineered features such as Fast Fourier Transform (FFT) frequency statistics were also calculated for each lead. This combination of clinically relevant and engineered features was used to train one classifier for each of the nine challenge diagnoses (AF, I-AVB, LBBB, Normal, PAC, PVC, RBBB, STD, STE) using the random forest algorithm, allowing for multiple diagnoses per recording. The random forest algorithm also allows for a high level of model interpretability which is essential for clinical translation. Each classifier reported a score from 0 to 1, with a decision threshold at 0.5. Five-fold cross validation was used to test the accuracy of each classifier, with recordings organized such that each fold contained the same proportion of recordings for each diagnosis. Recordings with multiple diagnoses were also equally split between the folds.

Results:     

The preliminary model resulted in a cross validated F-score of 0.309 and G-score of 0.176, with a geometric mean of 0.233. Using this model, the challenge submission F-score was 0.227 and G-score was 0.127, with a geometric mean of 0.170.