Assessing Intrinsic Cognitive Load in Virtual Reality Using Eye-Tracking: A Data-Driven Approach with a Novel Dataset

Understanding and assessing cognitive load (CL) in immersive virtual reality (VR) environments is crucial for the development of adaptive training systems and cognitive evaluation tools. This study introduces a non-intrusive eye-tracking-based framework to assess intrinsic cognitive load during structured VR tasks targeting memory, reasoning, and attention. These tasks were designed with progressively increasing complexity to induce varying levels of mental effort. To model the relationship between gaze behavior and task performance (used here as a proxy for cognitive load), we collected and analyzed eye-tracking data from users engaged in these tasks. Both feature extraction (including fixations, saccades, blinks, and gaze temporal dynamics) and deep learning on raw time-series data were employed. Among the methods evaluated, the best performance was achieved by a Random Forest classifier trained on extracted features with feature selection based on importance scores, achieving a test accuracy of 82.4%. A Hybrid CNN-LSTM model using raw sequential eye-tracking data also performed competitively, with a test accuracy of 81.6%. This work demonstrates the effectiveness of gaze-based behavioral features for real-time, scalable cognitive load assessment in VR and offers a novel dataset combining eye-tracking and performance data, addressing the lack of public benchmarks in this domain.