Identification of mental states in patients based on electrical activity is of interest for clinical neuroscience. While the problem has been successfully addressed for individual patients with multiple (up to 50) states identifiable for the purposes of communication, addressing the problem in a general sense is more difficult, as it requires understanding of the neural bases for successful classification. We used a range of well-established machine learning methods on a two-state classification problem based on EEG data with high success rate (>98%), and then applied Sparse Optimal Scoring (SOS) to reduce the dimensionality of the features, and improve model interpretability for generating an ‘explainable ML’ understanding of the basis for the successful classification. 24-channel EEG data from Deaf signers watching sign language videos and the same videos in reverse (non-comprehensible, but identical in spatiotemporal features) was used to calculate coherence between optical flow in the stimuli and EEG neural response (per video, per participant) using canonical component analysis. Peak correlations for binned frequencies were used as input parameter to machine learning algorithms. Two ensemble classifiers (AdaBoost and Random Forest) achieved 100 percent out-of-sample prediction accuracy on hold-out dataset for the whole brain. Sparse Optimal Scoring (SOS) was then applied to the coherence data to reduce the dimensionality of the features and improve model interpretability. SOS with elastic-net penalty resulted in out-of-sample classification accuracy of 98.89%. The sparsity patterns from the trials using 1 Hz bins consistently indicated frequencies between 0.2-1 Hz were primarily used in the classification. We find that successful classification relies on low (<2 Hz) frequencies of EEG coherence to the input signa. This indicates that that the hallmark of communicatively successful brain states is predictive processing for incoming signal.