Cognitive Load Classification using Functional Near-Infrared Spectroscopy

This thesis investigates the classification of cognitive load using functional near-infrared spectroscopy (fNIRS) signals recorded during an N-back working memory task. The study introduces a novel short-channel correction layer designed to suppress superficial physiological noise adaptively, addressing limitations of traditional General Linear Model (GLM) based regression. A single participant dataset comprising 69 validated sessions was analyzed using both conventional machine learning and deep learning approaches. Traditional classifiers: Linear Discriminant Analysis (LDA), Support Vector Machines (SVM), Random Forests, and Gradient Boosting were first evaluated using statistical features (mean, variance, peak, and slope). Among these, Gradient Boosting achieved the highest accuracy (55.6%), indicating that temporal features such as slope and peak were more discriminative than static measures. Building upon this baseline, two deep learning architectures: LSTM → FC×3 and CNN → LSTM → FC×2 were trained under four preprocessing conditions: baseline, short-channel correction, GLM correction, and combined SS + GLM correction. The CNN–LSTM hybrid achieved the best performance, with mean accuracy reaching 62.46% under the combined correction condition. Further architectural analyses showed that increasing LSTM depth reduced performance (55.9% → 52.1%), while removing the CNN layer caused a substantial drop (62.3% → 56.4%), confirming the importance of convolutional feature extraction. Expanding the number of fully connected layers modestly improved performance, indicating benefits from deeper nonlinear transformations. Overall, results demonstrate that combining short-channel correction with GLM regression and hybrid spatial–temporal deep learning architectures enhances fNIRS based cognitive load classification. These findings highlight the potential of adaptive noise correction and deep learning frameworks to improve the robustness and interpretability of fNIRS systems for real-world cognitive monitoring.