Evaluating the DoC-Forest tool for Classifying the State of Consciousness in a Completely Locked-In Syndrome Patient

This paper shows how the DoC-Forest method can be used for classifying the state of consciousness in a Completely Locked-In Syndrome (CLIS) patient based on electrophysiological measurements of the patients’ brain activity. The DoC-Forest is a Machine Learning (ML) based method to classify whether patients with a Disorder of Consciousness (DoC) are either in the minimally conscious state (MCS) or have the unresponsive wakefulness syndrome (UWS). As labeled training data is unavailable for CLIS patients, Stereoelectroencephalography (sEEG) data obtained from patients undergoing anesthesia was used to train an ensemble of decision trees to classify the state of consciousness. The model achieves a predictive area under the curve (AUC) of 0.73 on out-of-sample data during training. The most important features for training the model on the anesthesia dataset were determined, and an explainable ML algorithm was used to quantify how features contributed to forming the predictions for the CLIS patient. For the CLIS dataset comprising ECoG measurements from one CLIS patient, there was a time period where the CLIS patient clearly was conscious. Based on the predictions of the model trained on the anesthesia dataset, this period was clearly identifiable, indicating that despite the differences in the underlying conditions (CLIS and anesthesia) and the different measurement methods (sEEG and ECoG) the model was able to extract robust patterns predictive of the state of consciousness.