The PhysioNet 2021 challenge asks participants to develop automated techniques for classifying cardiac abnormalities (CA) from both 12-lead electrocardiogram (ECG) and reduced-lead settings. We investigated on the feasibility of applying Automated Machine Learning (AutoML) approaches to build ECG classifiers.
Standard ECG preprocessing was applied beforehand to the ECG (filtering and resampling). Three different AutoML frameworks were executed on the 88,000+ ECGs made available by the challenge organizers. The optimal combination of preprocessing and ML algorithms were found by the AutoML frameworks. We finally assessed the frameworks' classification performance, the effect of the number of employed leads, and the effect of extending the frameworks training time.
The classifiers proposed by our team "BiSP Lab" obtained a challenge score up to 0.35 on the private test set. The AutoML frameworks showed comparable performance. The worst score was obtained on the 12-lead system, while the best on the 6-lead one. Significantly extending the training time seemed to not improve the test score.
AutoML frameworks showed promising performance on the private test set, suggesting their potential for classification of CA. Future works are towards testing further AutoML approaches, and better determining the impact of the available training time on the classification performance.