Cognitive impairment assessment using eye-tracking: multilevel saccade paradigms with differential analysis and attention-based neural networks

The accurate assessment of cognitive impairment plays a vital role in more targeted treatments for Dementia. Eye movement analysis is a non-invasive and objective method that offers fine-grained insight into cognitive functioning, complementing conventional screening tools. However, single-task eye-tracking paradigms and simplistic analysis methods limit the potential for comprehensive and fine-grained assessment of cognitive impairment. To address this limitation, we propose a multilevel saccade paradigm combined with differential analysis and an attention-based neural network to enhance eye-tracking-based cognitive impairment assessment. Approach. Firstly, a set of saccade-based paradigms with graded difficulty levels is developed, including prosaccade, antisaccade, and random pro-antisaccade paradigms. Each paradigm incorporates eye movement assessments in both horizontal and vertical directions. Secondly, we recruit 90 subjects for eye-tracking assessments to build a large-scale dataset. The subjects consisted of 36 healthy young controls, 15 healthy elderly controls, 23 individuals with mild cognitive impairment, and 16 individuals with dementia. Each subject completed the Montreal Cognitive Assessment (MoCA). Third, the Mann-Whitney U test is employed to identify eye movement features that show significant differences across the four groups. Correlation analysis with MoCA scores further validated the effectiveness of these eye movement features in distinguishing cognitive impairment. Finally, XGBoost is employed to perform classification and to validate the effectiveness of the eye movement feature selection scheme derived from the difficulty-graded saccade paradigms. An attention-based neural network is also integrated to enhance classification accuracy and improve feature selection by identifying the most informative eye movement features. Main results. The model achieved an area under the receiver operating characteristic curve (AUC) of 0.94, a classification accuracy of 0.80, and a matthews correlation coefficient (MCC) of 0.73. Among all features extracted from the different saccade paradigms, the time to first correct AOI and saccade latency parameters from the random pro-antisaccade paradigm demonstrate the highest contribution to classification performance. Significance. By integrating graded saccade paradigms with statistical analysis and attention neural network, this study enhances the granularity and accuracy of eye-tracking-based cognitive assessment, offering a scalable and non-invasive tool for early detection and monitoring of cognitive decline.