Multimodal EEG and explainable machine learning characterize neuroticism-related neurodynamic heterogeneity in mild...

As a personality trait associated with emotional instability and negative emotional tendencies, neuroticism is considered to be closely related to the risk of cognitive decline in old age. However, the neurodynamic abnormalities associated with elevated neuroticism among individuals with MCI remain unclear.

Cognitive screening based in the community was performed in Zhumadian, Henan Province, and Xingtai, Hebei Province, China. Based on converted scores on the Neuroticism subscale of the Eysenck Personality Questionnaire, individuals with MCI were categorized into high-neuroticism MCI (HN-MCI) and non-elevated neuroticism MCI (NEN-MCI) groups. After EEG recruitment, behavioral-task completion, EEG quality control, and frequency matching by sex, age, and years of education, 53 participants with HN-MCI and 53 with NEN-MCI were included. All participants engaged in a modified visuospatial working memory test during the recording of 64-channel EEG data. Behavioral measures, event-related potentials, time-frequency power, phase-amplitude coupling, and dwPLI-based functional connectivity features were extracted. A stacking ensemble learning model integrated with SHAP-based interpretability analysis was employed for multimodal feature fusion and classification.

In comparison to the NEN-MCI group, the HN-MCI group exhibited lower Montreal Cognitive Assessment (MoCA) total and visuospatial/executive scores, prolonged reaction times, and elevated inverse efficiency scores. EEG analyses indicated increased P2 and P3 amplitudes, reduced N2 amplitude, increased theta activity, enhanced theta-to-low-alpha phase-amplitude coupling, and reinforced theta-band functional connectivity in the HN-MCI group. The stacking model differentiated the two groups with an AUC of 0.858, and permutation testing indicated that model performance was significantly above chance (permutation p < 0.001). SHAP analysis identified parietal P3 amplitude as the most influential feature.

Individuals with HN-MCI exhibit multilevel neurodynamic abnormalities across temporal, spectral, cross-frequency, and network domains during visuospatial working memory processing, accompanied by impaired visuospatial/executive function and reduced behavioral efficiency. Neuroticism-related emotional traits may contribute to a sustained high-load and low-efficiency neural processing pattern in individuals with MCI. Multimodal EEG combined with explainable machine learning provides a useful framework for characterizing personality-related neurodynamic heterogeneity in MCI.