One must be aware of the black-box problem by applying machine learning models to analyze high-dimensional neuroimaging data. It is due to a lack of understanding of the internal algorithms or the input features upon which most models make decisions despite outstanding performance in classification, pattern recognition, and prediction. Here, we approach the fundamentally high-dimensional problem of classifying cognitive brain states based on functional connectivity by selecting and interpreting the most relevant input features. Specifically, we consider the alterations in the cortical synchrony under a prolonged cognitive load. Our study highlights the advances of this machine learning method in building a robust classification model and percept-related prestimulus connectivity changes over the conventional trial-averaged statistical analysis.