An ensemble of Xgboost models for detecting disorders of consciousness in brain injuries through EEG connectivity

Disorders of consciousness (DoC) happen frequently in brain injuries. Automated DoC detection is desirable because conventional clinical examinations, typically behavioural assessments, consume significant resources of manpower and time. Extracting knowledge from electroencephalogram (EEG) signals and then employing machine learning to classify patients are an attractive way to achieve automatic detection of DoC. However, an EEG dataset from brain injuries always has much more subjects without DoC than the subjects with DoC. Because of the heavy imbalance of the dataset, even if the re-sampling technique is used, it is still challenging to achieve good classification performance from, e.g., eXtreme Gradient Boosting (Xgboost), one of the most effective machine learning methods. This paper develops an ensemble of multiple Xgboost models (EoXgboost) to improve the machine learning performance for DoC detection in the presence of severe class distribution skews. We also explore effective connectivity indices to distinguish brain injuries with and without DoC. A new combined connectivity measure is shown to be effective in detecting DoC in a dataset of 648 brain injuries. Our classification results show that our EoXgboost classifier with the new combined connectivity index has achieved the best classification performance than the single connectivity indices. Our EoXgboost has detected DoC in brain injuries with the accuracy of 99.07%, AUC of 98.74%, specificity of 99.77%, and sensitivity of 97.71%. This demonstrates that our EoXgboost is a promising tool for DoC detection in brain injuries.