Hyperlet Transform-Based Cognitive Load Assessment Using Single-Channel fNIRS Signals

Abstract

Cognitive load assessment using functional nearinfrared spectroscopy (fNIRS) has gained increasing attention due to its non-invasive and high spatial resolution nature. However, reliable classification remains challenging due to non-stationary and nonlinear nature of haemodynamic signals. Conventional time-frequency representation methods often fails to capture subtle amplitude variations from haemodynamic responses due to limited joint time-frequency resolutions. In this study, an attempt has been made to characterize and estimate cognitive load using single-channel fNIRS and Hyperlet Transform (HLT). The framework integrates preprocessing, HLT based time-frequency decomposition, multidomain feature extraction, and optimal feature selection within a Leave-One-Subject-Out (LOSO) folds to identify robust biomarkers. The extracted features are applied to five ML classifiers namely SVM-RBF, KNN, Naive Bayes, AdaBoost, and Random Forest. Among the evaluated classifiers, Support Vector Machine with a Radial Basis Function kernel achieved highest performance with an accuracy of 73.33% for 3-class tasks, followed by 86.67% and 90.00% for Baseline vs Low and Baseline vs High Cognitive Load tasks. Further, HLT outperformed conventional time-frequency methods, which emphasizes that HLT derived features effectively capture cognitive workload related hemodynamic changes and provides strong potential for development of real-time deployable solutions for next-generation affective computing systems.