Key points are not available for this paper at this time.
Background We investigated differences in memory-guided saccades (MGS) and electroencephalographic (EEG) spectral ratios between cognitively normal older adults and patients with mild cognitive impairment (MCI). A multimodal predictive model was developed to evaluate its potential for early MCI identification. Methods We enrolled 60 participants (37 MCI, 23 controls) who underwent cognitive assessments, 32-channel resting-state EEG, and MGS testing. EEG spectral ratios, including the Delta-to-Alpha Ratio (DAR), Theta-to-Alpha Ratio (TAR), and (Delta+Theta)/(Alpha+Beta) Ratio (DTABR), were extracted along with MGS parameters (latency, accuracy, gain). To prevent data leakage, 21 multimodal features were evaluated using a zero-leakage least absolute shrinkage and selection operator (LASSO) pipeline. A multivariable predictive model was subsequently constructed using Firth’s penalized logistic regression, explicitly adjusting for demographic covariates, and internally validated with 1,000 bootstrap resamples. Results MCI patients exhibited significantly prolonged bilateral saccadic latencies, decreased accuracy, and widespread cortical spectral slowing (elevated DAR, TAR, and DTABR across all regions) (all P 0.05). The LASSO model identified a seven-feature cross-modal diagnostic panel. Following Firth’s penalization, rightward saccadic accuracy maintained its significance as an independent protective factor. The combined nomogram achieved an apparent AUC of 0.994 and a bootstrap optimism-corrected AUC of 0.975, demonstrating favorable calibration and promising net clinical benefit. Conclusion MCI patients demonstrate coupled oculomotor and electrophysiological abnormalities, reflecting impaired frontoparietal-sensorimotor network efficiency. Our rigorously penalized multimodal model yields robust internal diagnostic performance, providing a promising exploratory proof-of-concept for the early screening of “cognitive-motor decoupling” in clinical neurology, though large-scale external validation is warranted.
Zhao et al. (Thu,) studied this question.