This study assessed the effectiveness of three digital screening tools in detecting cognitive impairment (CI) in a large cohort of community-dwelling elderly individuals and investigated the relationship between key digital features and plasma p-tau217 levels.

This community-based cohort study included 1,083 participants aged 65 years or older, with 337 diagnosed with CI and 746 classified as normal controls (NC). We utilized two screening approaches: traditional methods (AD8, MMSE scale, and APOE genotyping) and digital tools (drawing, gait, and eye tracking). LightGBM-based machine learning models were developed for each digital screening tool and their combination, and their performance was evaluated. The correlation between key digital features and plasma p-tau217 levels was analyzed as well.

A total of 21 drawing, 71 gait, and 35 eye-tracking parameters showed significant differences between the two groups (all p < 0.05). The area under the curve (AUC) values for the drawing, gait, and eye-tracking models in distinguishing CI from NC were 0.860, 0.848, and 0.895, respectively. The combination of eye-tracking and drawing achieved the highest classification effectiveness, with an AUC of 0.958, and accuracy, sensitivity, and specificity all exceeded 85%. The fusion model achieved an AUC of 0.928 in distinguishing mild cognitive impairment (MCI) from NC. Additionally, several digital features (including two drawing, ten gait, and one eye-tracking parameters) were significantly correlated with plasma p-tau217 levels (all |r| > 0.3, p < 0.001).

Digital screening tools offer objective, accurate, and efficient alternatives for detecting CI in community settings, with the fusion of drawing and eye-tracking providing the best performance (AUC = 0.958).