Consciousness and its absence can be assessed and quantified through different techniques. Sleep-stage classification via electroencephalography (EEG) signal analysis is the most common and effective method for this purpose. Recently, the research moved from a spectral investigation of the EEG to a complexity study. Entropy features showed to be more robust biomarkers of consciousness than frequency-based measures. To further contribute to this discussion, we utilized a Deep Learning (DL) derived function, Gradient-weighted Class Activation Mapping (Grad-CAM), to predict if this generalizability is found also in the DL's unbiased feature selection. To do so, we exploited a semi-automatic Machine Learning (ML) classification pipeline to compare the performances obtained with the original EEG and the Grad-CAM.