Enhancing Mental Workload Classification in Neuroscience: Deep Learning Approaches

Abstract This study aims to classify mental workload levels from n-back cognitive data by employing a diverse set of deep learning approaches capable of accommodating both dense and sparse features. The n-back task paradigm provides rich temporal data, capturing information on working memory and cognitive demand from the same subjects over time across different n-back conditions. By integrating deep learning techniques that leverage both dense and sparse features, this research introduces novel perspectives tailored to the data structure of the n-back task. Our findings highlight the effectiveness of the extreme Deep Factorization Machine model with stratified 5-fold cross-validation. Compared to the baseline model for the 0- vs 1-back classification task, this approach achieved significant improvements: 67.50% in accuracy, 68.74% in sensitivity, 66.24% in specificity, and 68.48% in F1-Score among the models in which dense features are the combinations of hemodynamic measures and experimental variable, and subject is considered as a sparse feature. Additionally, employing Principal Component Analysis (PCA) resulted in significant enhancements in performance metrics compared to the baseline Logistic Regression model. Specifically, the utilization of PCA led to remarkable improvements of 53.03% in accuracy, 98.03% in sensitivity, 24.14% in specificity, and 70.37% in F1-Score, as observed in the classification of the 0- vs 1-back condition when using the xDeepFM model. These results underscore the utility of deep learning methods in accurately discerning mental workload levels from complex n-back cognitive science data, offering valuable insights into cognitive functioning and workload assessment.