Electroacupuncture modulates Electroencephalographic microstate dynamics to alleviate chronic insomnia: A machine learning approach for predicting individual treatment response

Enqi Liu^1,2,3Chi Wang⁴Xiaoqiu Wang¹Kai Liu¹Shan Qin³Liyu Lin¹Juan Li¹Min Xu¹Liu Chengyong³Huangan Wu⁴Wenzhong Wu^1*

• ¹Nanjing University of Chinese Medicine, Nanjing, China
• ²Nanjing University of Chinese Medicine - Hanzhongmen Campus, Nanjing, China
• ³Nanjing Hospital of Chinese Medicine Affiliated to Nanjing University of Chinese Medicine, Nanjing, China
• ⁴Shanghai University of Traditional Chinese Medicine, Shanghai, China

Background: Chronic insomnia (CI) is associated with dysregulation of brain network dynamics, and patient response to electroacupuncture (EA) treatment varies. This study aimed to investigate the characteristics of electroencephalographic(EEG) microstates in patients with CI, analyze changes in microstate parameters before and after EA treatment, and explore the potential application of machine learning (ML) models based on baseline microstate features for predicting treatment response. Methods: We enrolled 41 CI patients and 19 healthy controls (HC). Baseline resting-state EEG was recorded, and microstate parameters (classes A–D) were analyzed. CI patients underwent 4-week EA treatment. Six clinical scales—including the Pittsburgh Sleep Quality Index (PSQI) and Hamilton Depression Scale, and microstate dynamics were compared pre-and post-treatment. Treatment response was defined as ≥50% PSQI reduction. Multi-stage feature selection and eight ML algorithms were used to build the prediction model. Results: At baseline, CI patients showed differences in some temporal metrics of microstates B, A, and C compared to HC. After EA, all clinical scores improved significantly (p < 0.001). Coverage_B and Duration_B, as well as Occurrence_C, increased, and multiple transition probabilities were regulated—particularly, microstate B temporal indicators normalized to HC levels. In the exploratory ML modeling, RF performed best (AUC = 0.849). "Duration_A", "OrgTM_D→B", and "OrgTM_C→B" were the top positive predictors, while "Occurrence_C" and "Duration_B" were negative predictors. Conclusion: This study found that EA treatment was associated with improved clinical scores and alterations in some EEG microstate parameters in patients with CI. In this exploratory analysis with a limited sample size, baseline microstate features showed preliminary potential for predicting treatment response, though further validation in larger cohorts is needed. These findings may provide a reference for future research on neurophysiological predictors and the development of individualized treatment strategies.