Abstract
OBJECTIVES: Spinal cord stimulation (SCS) is a commonly used treatment for refractory zoster-associated pain (ZAP). The conventional open-loop regulation mode mainly relies on patients' subjective feedback for manual parameter adjustment, whereas the clinical application of closed-loop SCS regulation remains immature and requires further optimization. Previous studies have shown that SCS can influence electroencephalography (EEG) functional connectivity. This study aims to investigate whether a support vector machine (SVM) based on EEG functional connectivity features can effectively identify different neuromodulation states of SCS. METHODS: A total of 21 patients with ZAP who underwent SCS treatment were retrospectively included. Visual Analogue Scale (VAS) scores were recorded before SCS implantation, at discharge, and at 1, 3, and 6 months postoperatively. EEG data lasting 5-10 minutes were collected under SCS-on and SCS-off conditions. The EEG data were randomly divided into a training set and a test set at a ratio of 7∶3. Pearson correlation coefficient (PCC) and coherence (COH) between EEG feature channels were extracted to construct an SVM-based classification model. Model performance in distinguishing SCS-on and SCS-off EEG states was evaluated using accuracy, sensitivity, specificity, and the area under the receiver operating characteristic curve (AUC). Shapley additive explanations (SHAP) were used to interpret the SVM model by analyzing the contribution and direction of each EEG feature to the model prediction. RESULTS: SCS effectively alleviated pain in patients with ZAP. Compared with baseline, VAS scores were significantly reduced at discharge and at 1, 3, and 6 months after SCS treatment (all P<0.001). The classification performance of SVM with different kernel functions for identifying SCS neuromodulation states showed that the radial basis function kernel SVM achieved an accuracy of 0.912, sensitivity of 0.927, specificity of 0.893, and AUC of 0.972, whereas the linear kernel SVM achieved an accuracy of 0.677, sensitivity of 0.714, specificity of 0.665, and AUC of 0.744. SHAP analysis showed that the top 20 most influential features contributing to the classification results were all COH features. In the beta frequency band, the COH between electrode pairs O(2)-F(8) and FP(2)-F(7) contributed most strongly to the model prediction, and increases in these values were positively associated with the SCS-on state. In contrast, the COH of the P(4)-P(8) electrode pair in the theta frequency band was closely associated with the SCS-off state. CONCLUSIONS: The SVM model based on EEG functional connectivity features can effectively identify different neuromodulation states of SCS in patients with ZAP. These findings suggest that SCS may induce alterations in brain functional connectivity and provide a basis for future studies on EEG-based closed-loop SCS regulation.