Multimodal Evaluation of Cognitive Reserve in a Healthy Sample Using Eye-Tracking, Performance Features, and Machine Learning Models

The paper presents a machine learning-based approach to assess cognitive workload and cognitive reserve using a Digit Symbol Substitution Test (DSST) framework combined with eye-tracking data. A total of 61 participants completed three progressively challenging DSST trials under time pressure, with eye movements, pupil dynamics, and performance metricsrecorded by an eye tracker. The DSST was adapted to represent four distinct levels of cognitive workload by varying thecomplexity of symbols and task duration. A classification analysis using XGBoost, Random Forest, and SVM models achieveda precision of up to 75% and an AUC of 81%. NASA-TLX results confirmed a progressive increase in perceived workload, particularly in mental effort, and supported the test’s sensitivity to task complexity. Linear regression and hierarchical clustering analysis further revealed general trends in feature behavior across workload levels, as well as individual variability suggestive of cognitive reserve in a subset of participants. These individuals maintained stable or improved performance under highercognitive load despite physiological indicators of strain. Overall, this work demonstrates that combining eye-tracking data with machine learning can effectively classify cognitive workload levels and uncover inter-individual differences in cognitive resilience, offering valuable insights for future cognitive assessment and mental fatigue monitoring frameworks.