Diagnosis of disorders of consciousness using nonlinear feature derived EEG topographic maps via deep learning.

This study explored the value of nonlinear features extracted from EEG signals to facilitate the assessment of patients with disorders of consciousness (DOC) with limited communication capacity. We utilized a dataset comprising 104 participants, 56 with vegetative state (VS)/unresponsive wakefulness syndrome (UWS) and 48 in a minimally conscious state (MCS). For each participant, we computed channel-wise approximate entropy (ApEn) from EEG time-series data using a sliding window approach under two experimental paradigms: resting state and preferred music stimulation. These nonlinear measures were then spatially interpolated to generate topographical maps. Both resting state and preferred music stimulation data were processed as 1-second epochs using identical convolutional neural networks (CNN) architectures. The classification performance and validity of the CNN were compared against support vector machine (SVM) and generalized regression neural network (GRNN) models. ApEn in the resting state and under stimulation with preferred music correlated with the Coma Recovery Scale-Revised scores in patients with DOC, showing varied regional responses. Notably, the CNNs resulted in a positive diagnostic performance with an accuracy of 90.00% and an AUC of 0.902. The CNN was better than the SVM and GRNN in differentiating between the VS/UWS and MCS states. This study offers a convenient and accurate method for detecting awareness in patients with VS/UWS and MCS using ApEn features in the resting state and under preferred music stimulation using deep learning.