Classification of 12-lead ECGs Using Gradient Boosting on Features Acquired With Domain-Specific and Domain-Agnostic Methods
This year, the objective of the PhysioNet/Computing in Cardiology challenge was the classification of 12-lead electrocardiograms (ECG). The approach presented in this paper consists of two parts, feature extraction and classification. The extracted features can be separated into domain-specific and domain-agnostic features, where domain-specific features are based on known ECG processing methods such as QRS-detectors. Domain-agnostic features are generated by wavelet transforms that take the raw 12-lead ECG as input. Additionally, a novel beat-to-beat correlation analysis is proposed to identify arrhythmia occurring among other healthy beats. These features are then combined and classified by gradient-boosted trees implemented in Python. To account for the complexity of the multi-label and multi-class problem definition, a One-vs-Rest scheme is utilized, where distinct classifiers for each class determine whether a sample belongs to said class. The resulting imbalance in training sets for each classifier was compensated for by giving the positive samples a higher weight. The classifiers were trained using the XGBoost gradient boosting system. The proposed classification scheme of the team “desafinado” received a score of 0.576 on the validation dataset and a score of 0.233 on the test set of the challenge (rank 19 of 41).
Paper
Full text
Classification of 12-lead ECGs Using Gradient Boosting on Features Acquired With Domain-Specific and Domain-Agnostic Methods
Semantic Scholar · Medicine · 2020
Abstract
This year, the objective of the PhysioNet/Computing in Cardiology challenge was the classification of 12-lead electrocardiograms (ECG). The approach presented in this paper consists of two parts, feature extraction and classification. The extracted features can be separated into domain-specific and domain-agnostic features, where domain-specific features are based on known ECG processing methods such as QRS-detectors. Domain-agnostic features are generated by wavelet transforms that take the raw 12-lead ECG as input. Additionally, a novel beat-to-beat correlation analysis is proposed to identify arrhythmia occurring among other healthy beats. These features are then combined and classified by gradient-boosted trees implemented in Python. To account for the complexity of the multi-label and multi-class problem definition, a One-vs-Rest scheme is utilized, where distinct classifiers for each class determine whether a sample belongs to said class. The resulting imbalance in training sets for each classifier was compensated for by giving the positive samples a higher weight. The classifiers were trained using the XGBoost gradient boosting system. The proposed classification scheme of the team “desafinado” received a score of 0.576 on the validation dataset and a score of 0.233 on the test set of the challenge (rank 19 of 41).