Classification of 12-lead ECG with an Ensemble Machine Learning Approach

Matteo Bodini1, Massimo W Rivolta2, Roberto Sassi3
1Università degli Studi di Milano, 2Dipartimento di Informatica, Università degli Studi di Milano, 3Università degli Studi di Milano, Dipartimento di Informatica


Abstract

Aims: The PhysioNet 2020 Challenge focuses on the automatic classification of cardiac abnormali-ties in 12-lead ECG signals. In this work, we addressed the classification problem by means of an ensemble machine learning approach, by extracting average ECG morphologies and rhythm-related features.

Methods: We used only the data provided for the challenge. First, we applied standard ECG pre-processing to select good quality signals. Second, we extracted the average ECG morphology and rhythm-related features based on the kind of cardiac abnormality to be detected. For AF, I-AVB, LBBB, RBBB, STD, and STE signals, we computed multi-lead average ECG segments depending on the abnormality under scope, to obtain a feature vector representing the average morphology. For signals labeled with AF, PAC, and PVC, we computed rhythm-related features, e.g. the average RR interval, and its standard deviation. Third, we designed an ensemble machine learning model us-ing three fully connected neural networks for detecting AF, I-AVB, LBBB, RBBB, STD, and STE, and three random forests for AF, PAC, and PVC. The final classification was performed by combining the outcomes of each classifier using a heuristic. Data at disposal were randomly split with 70/30 ratio to obtain a training and validation set. Hyperparameters were set using a 10-fold cross vali-dation performed on the training set.

Results: The ensemble model achieved a validation accuracy ranging between 0.84 and 0.96, de-pending on the type of abnormality. The best instance submitted by our team “BiSP Lab” during the unofficial phase achieved a F_2 score of 0.521, G_2 score of 0.321, and a geometric mean of 0.409.

Conclusions: Our ensemble classifier showed potential for classification of cardiac abnormalities. Future investigations are towards the testing of different heuristics for merging the decisions of the ensemble, assessing the robustness on ECGs of lower quality, and the evaluation of different machine learning models.