Abstract
BACKGROUND: Accurate diagnosis and assessment of mild cognitive impairment (MCI) are essential. The efficacy of saccades in the detection of MCI lacks validation through large-scale clinical trials. METHODS: All eligible participants underwent saccadic assessment in four tasks and a cognitive assessment. MCI diagnoses were made on the basis of clinical indicators and MRI by experienced physicians. The physicians were blinded to the saccade experiments, and the operators of the saccade experiments were blind to the diagnosis of physicians. The classification models based on machine learning were constructed for assessing the diagnostic accuracy of MCI based on saccadic parameters. RESULTS: Of the 559 residents who consented to participate, 383 (153 with MCI and 230 controls) were completely assessed. The classification model trained by saccadic parameters achieved high accuracy in dissociating MCI and control with an area under the curve (AUROC) of 0.945 (95% CI, 0.924-0.964), sensitivity of 0.824 (95% CI, 0.769-0.886) and specificity of 0.904 (95% CI, 0.867-0.935). The parameters of the memory-guided and antisaccade tasks demonstrated better diagnostic efficacy. The saccade model also exhibited a good diagnostic value in patients with borderline cognition, being defined by the score of the Montreal Cognitive Assessment (MoCA). When the borderline cognition was defined as 23-27 of the MoCA score, the diagnosing accuracy of mild cognitive impairment based on saccadic parameters resulted in AUROC of 0.911 (95% CI, 0.836-0.972), sensitivity of 0.929 (95% CI, 0.762-1.000) and specificity of 0.796 (95% CI, 0.718-0.863). CONCLUSIONS: Saccades can distinguish MCI from controls with great accuracy, offering a sensitive and objective diagnostic aid of MCI, especially in participants with borderline cognition.