Development and validation of a machine learning model integrating clinical and structural brain morphometric features for predicting postherpetic neuralgia: a multicenter cohort study with retrospective development and prospective external validation in China

Postherpetic neuralgia (PHN) is a common and difficult-to-treat complication of herpes zoster (HZ). Early identification of patients at high risk may facilitate risk stratification and individualized intervention. This study aimed to integrate clinical variables with three-dimensional T1-weighted structural magnetic resonance imaging (sMRI) features to develop an interpretable model for predicting PHN. This multicenter cohort study retrospectively included 257 hospitalized patients with acute or subacute herpetic neuralgia between February 2023 and October 2025 and prospectively enrolled 50 hospitalized patients between December 2025 and March 2026 for external validation. The development cohort was randomly divided into training and internal validation sets at a 7:3 ratio. Clinical predictors were selected using univariable and multivariable logistic regression, whereas sMRI features were selected using Spearman correlation analysis, the Mann-Whitney U test, and least absolute shrinkage and selection operator (LASSO) regression. Five algorithms were used to develop clinical, sMRI, and combined clinical-sMRI (CR) models. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), Brier score, calibration curves, and decision curve analysis. SHapley Additive exPlanations (SHAP) were used to interpret the models, and subgroup and stratified analyses were performed. Among 303 patients, 163 (53.8%) developed PHN. Four clinical features—Numerical Rating Scale (NRS) score, disease phase, time to antiviral treatment initiation, and time to nerve block initiation—and five brain morphometric features involving the right anterior cingulate and paracingulate gyri, right cuneus, right fusiform gyrus, left superior parietal gyrus, and right hippocampus were retained. The CR-SVM model showed consistent performance, with AUCs of 0.801 (95% CI, 0.735–0.867) in the training set, 0.767 (95% CI, 0.657–0.876) in the internal validation set, and 0.770 (95% CI, 0.633–0.906) in the external test set. Performance was more consistent across cohorts among patients in the acute phase, and discrimination was higher among patients with head and neck zoster. NRS scores of 4–6, the subacute phase, and right hippocampal gray matter volume > 0.005417 were associated with an increased risk of PHN, with adjusted risk ratios of 2.28, 1.34, and 1.44, respectively. Right cuneus cortical thickness > 1.816 mm was associated with a lower risk (adjusted risk ratio, 0.61; all P < 0.05). The model may require further recalibration in larger independent cohorts. Clinical and structural brain morphometric features provided complementary predictive information. The CR-SVM model identified patients at high risk of PHN with moderate discrimination and provided an interpretable approach to risk stratification, while offering neuroimaging clues to the central mechanisms underlying pain chronification. Not applicable.