Patients in a minimally conscious state (MCS) are characterized by behavioral signs of self- or environmental-awareness. EEG oscillations in MCS are known slowing down. Neuroelectrical modulations of MCS intervention – e.g., transcranial direct current stimulation (tDCS) – are assessed using EEG features, and related to clinical scores. However, these features are handcrafted in a non-patient-specific way. Conversely, interpretable and explainable artificial intelligence (IXAI) automatically extracts the most salient patient-specific features. In this pilot study, we use IXAI for tracking the EEG changes of 9 MCS patients following tDCS. Our approach is composed by an interpretable neural network (Sinc-ShallowNet) and an explanation technique (DeepLIFT). The network discriminates resting-state EEG before vs. after tDCS and learns interpretable spectral features; DeepLIFT quantifies the relevance of the frequency components. Patients with higher neurobehavioral improvements show a maximum relevance at high frequencies (high-alpha) and a minimum at low frequencies (delta); vice versa for patients with lower improvements. Our results corroborate the idea that tDCS could support MCS intervention, and that IXAI could be useful, prospectively, to design patient-specific markers tracking the effects of intervention.
