Automatic Diagnosis of Cardiac Disease from Twelve-lead and Reduced-lead ECGs using Multi-label Classification

Prathic Sundararajan, Kevin Moses, Cristhian Potes, Saman Parvaneh
Edwards Lifesciences


Background: ECG is an essential tool for the clinical diagnosis of multiple cardiac diseases. The 2021 PhysioNet/CinC Challenge aims to develop algorithms to classify cardiac abnormalities from Twelve-lead and Reduced-lead ECGs. In this article, RandOm Convolutional KErnel Transform (ROCKET) features with a multi-label classification using XGBoost were used for this challenge. Method: About 40% of the training dataset, consisting of 32,200 ECG records, had more than one scored diagnosis. Therefore, multi-label classification was selected for ECG classification. Iterative stratification is used to split the training data into ten folds. Eight and two folds were selected randomly as in-house training and test set, respectively. ROCKET features were extracted from all available leads for each record. The label powerset approach was used to transform a multi-label problem into a multi-class problem—one multi-class classifier XGBoost trained on all unique label combinations found in the training data. Python was used for feature extraction and model development. Results: Our classifiers received scores of 0.504, 0.466, 0.459, 0.458, and 0.438 for the 12-lead, 6-lead, 4-lead, 3-lead, and 2-lead versions on the in-house test set with the challenge evaluation metric (10 seconds ECG). Unfortunately, we could not get scores for our submissions on the challenge hidden validation and test set. Discussion:
We proposed a multilabel XGBoost for classifying cardiac abnormalities from twelve-lead and reduced-lead ECGs using RandOm Convolutional KErnel Transform (ROCKET) with 10,000 kernels for feature extraction. Promising results on the in-house test set on the public training set indicate the power of ROCKET for feature extraction.