Postherpetic neuralgia risk prediction in hospitalised patients with herpes zoster based on MIMIC-IV.

Postherpetic neuralgia (PHN) is a disabling complication of herpes zoster (HZ). Early identification of high-risk patients may enable targeted prevention and follow-up. We developed and compared machine-learning models for PHN prediction using the Medical Information Mart for Intensive Care IV database and identified key predictors. This retrospective cohort study used Medical Information Mart for Intensive Care IV v2.2 (2008-2019). Adult inpatients with HZ were identified via International Classification of Diseases-9/10 codes (053.x or B02.x). The outcome was PHN (binary). Candidate predictors included demographics, pain phenotypes, comorbidities, laboratory markers, and inpatient medications; an antiviral initiation delay  ≥ 3 days (AVdelay3d) was derived. Data were split into stratified training/validation sets (70/30). Five-fold cross-validation with Bayesian optimization was applied for categorical boosting, light gradient-boosting machine, extreme gradient boosting (XGBoost), and adaptive boosting; least absolute shrinkage and selection operator logistic regression was tuned via cross-validation. Discrimination was assessed by area under the receiver operating characteristic curve; the optimal probability threshold was determined by maximizing the Youden index. Feature selection used Boruta and least absolute shrinkage and selection operator, supplemented by feature importance from the best-performing model. All models were validated internally using the same dataset, and no external validation was performed. Of 383 inpatients with HZ, 81 (21.1%) developed PHN. XGBoost achieved the best discrimination (area under the receiver operating characteristic curve = 0.737; 95% confidence interval: 0.652-0.822). Baseline comparisons showed that PHN risk was associated with a higher maximum pain score, AVdelay3d, diabetes, amitriptyline use, a lower hematocrit level, and severe immunosuppression (all P < .05). In the Boruta analysis, the confirmed important features (in descending order) were D-dimer, hematocrit, creatinine, AVdelay3d, age, and platelet count. In the PHN prediction model, tree-based methods (e.g., XGBoost) can effectively estimate PHN risk and outperform linear models. These findings suggest that the pathogenesis of PHN may involve complex, nonlinear interactions. Baseline comparisons and feature-importance analyses further identified a set of core predictors for PHN, providing a reference for early identification of high-risk patients and for implementing risk-stratified management. However, the results should be interpreted cautiously because only internal validation was performed, and external validation was lacking.