Binary vs. Multiclass Mental Workload Classification from fNIRS Data: A Deep Learning-Based Subject-Specific Analysis

Functional near-infrared spectroscopy (fNIRS) is a portable and non-invasive neuroimaging technique increasingly utilized for mental workload monitoring in cognitive and human-computer interaction studies. In this study, the effectiveness of time-series deep learning models for subject-specific classification of mental workload levels from fNIRS signals was evaluated. Three classification approaches, two binary and one multiclass, were applied using bandpass-filtered and segmented fNIRS data from nine participants selected based on age, gender, handedness, and previous n-back task experience. We used CNN, CNN-GRU, and CNN-LSTM architectures selected for their ability to capture spatial patterns and model temporal dependencies in multivariate sequences. All the models were trained and tested on normalized temporal segments using consistent hyperparameters. Our results revealed that the CNN-LSTM model achieved the highest performance in the binary classification, where label 0 was assigned as low workload and label 2 as high workload, reaching an average test accuracy of 92.35% ± 5.06, F1-score of 0.923 ± 0.051, and AUC of 0.981 ± 0.023.