Fig. 1. Overview of the study.

a Data description. The same participants participated in transcranial magnetic stimulation-induced electroencephalography (TMS–EEG) and resting-state electroencephalography (EEG) measurements. The sleep condition did not include resting-state EEG, and one participant under ketamine-mediated anesthesia was missing in resting-state EEG. b Schematic framework for determining the explainable consciousness indicator (ECI). In step 1, raw EEG signals were converted into a spatio-spectral or spatiotemporal 3D matrix. In step 2, the converted 3D feature was used on a convolutional neural network in the two components of consciousness: arousal and awareness. In each arousal and awareness state, the EEG data were trained as two classes (low versus high). For example, for awareness, rapid eye movement (REM) sleep with subjective experience (i.e., dreaming) and healthy wakefulness belong to the same class in terms of high awareness; however, for arousal, non-rapid eye movement (NREM) with no subjective experience and REM sleep with subjective experience belong to the same class in terms of low arousal. The output $p_i$ indicates the probability in the trial i of arousal and awareness. For the training and test phase, we used the leave-one participant-out approach as transfer learning. Therefore, the EEG data in the source pool were used for training and the data of target participants was predicted for arousal or awareness. The source pool contains data corresponding to the source domain except for the target participant. In step 3, the interclass probability for each arousal and awareness was averaged for calculating ECI in each session j. The averaged probability $C_j$ is ECI^aro and ECI^awa on the x- and y-axes, respectively. Therefore, we represented the 2D consciousness indicator for the two components of consciousness. In the final step, we checked which brain signals the model has learned and why it made such a decision using layer-wise relevance propagation (LRP). Through this step, we could interpret the proposed indicator. ECI^awa = ECI in awareness component; ECI^aro = ECI in arousal component.