Abstract
OBJECTIVE: This study aims to investigate the differences in electroencephalogram (EEG) microstates and power spectrum between patients with epilepsy (PWE) comorbid with (without) mild cognitive impairment (MCI) and to develop a machine learning model to predict the risk of MCI comorbidity in PWE. METHOD: Participants were classified into PWE comorbid with MCI (EPMCI) and PWE comorbid without MCI (EPNMCI). The microstate parameters and power spectral density (PSD) of both groups were compared. We combined different types of variables and constructed models using Support Vector Machine (SVM), Neural Network (NNET), Random Forest (RF), K-Nearest Neighbors (KNN), and Naive Bayes (NB). An ideal predictive model was selected to evaluate the risk of MCI comorbidity in PWE. RESULT: A total of 627 PWE were included in this study, of whom 106 had MCI and 521 did not. Significant differences were observed between the two groups of patients in microstates A, B, C, D, and PSD. Among various machine learning models and multiple variable groups, we selected the NNET model based on microstate variables as the optimal model. It demonstrated the second-highest ROCAUC value (0.93), the highest accuracy (0.89), the lowest standard error (0.11), and superior calibration metrics, including the highest discrimination index (D = 0.724), the lowest Brier score (0.084), and the smallest unreliability index (U = 0.006). Finally, we compared this model with the traditional MMSE decision curve analysis (DCA) and found that it exhibited a wider range of applicable thresholds and a greater overall net benefit, demonstrating enhanced clinical utility. CONCLUSION: Differences in EEG microstates analysis and spectral analysis provide evidence for the mechanisms and dynamic changes associated with epilepsy comorbid with MCI. The development of a predictive model offers guidance for the assessment of MCI in specific populations with epilepsy.