A multicenter prospective cohort study developing and validating a SIRI-based machine learning model and simplified...

Postherpetic neuralgia (PHN) is the most common and severe complication of herpes zoster (HZ), and early identification of high-risk patients remains a major clinical challenge. The systemic inflammatory response index (SIRI) is a novel inflammatory marker that has demonstrated prognostic value in multiple diseases, but its association with the risk of PHN development has not been reported to date. Existing PHN prediction models mostly rely on traditional regression methods with limited predictive performance, and lack clinically practical risk stratification tools.

After screening for eligibility and follow-up completion, 1135 patients from the Hangzhou Third People's Hospital and 226 patients from the Second Affiliated Hospital of Zhejiang Chinese Medical University were included in the final statistical analysis. 1135 enrolled patients were stratified and randomly split at a 3:1 ratio into a training set for model development and an internal test set for internal validation, while data from 226 enrolled patients served as the independent external validation set. A total of 36 candidate predictive factors were first screened via univariate Logistic regression (LR) and collinearity diagnosis, followed by Least Absolute Shrinkage and Selection Operator (LASSO) regression with 10-fold cross-validation to identify core predictive factors. Finally, the predictive performance of 8 machine learning algorithms was compared, and a simplified risk scoring table was constructed based on SHapley Additive exPlanations (SHAP) analysis.

A total of 1361 patients were finally included, with 511 (37.55%) developing PHN during follow-up. Multivariate binary LR confirmed that SIRI was an independent predictive factor for PHN (odds ratio [OR] = 1.448, 95% confidence interval [CI] 1.119-1.874, P = 0.005). LASSO regression identified 6 core predictive variables: age, SIRI, numerical rating scale (NRS) score, time to treatment, rash location, and neutrophil-to-albumin ratio (NAR). The eXtreme Gradient Boosting (XGBoost) model exhibited the optimal predictive performance, with area under the receiver operating characteristic curve (AUC) values of 0.889 (95% CI 0.868-0.910) in the training set, 0.857 (95% CI 0.813-0.901) in the internal test set, and 0.900 (95% CI 0.860-0.940) in the external validation set. SHAP analysis revealed that SIRI ranked second in feature contribution to the model (20.1%), and the simplified risk scoring table constructed based on SHAP values achieved an AUC of 0.904 (95% CI 0.867-0.941) in the external validation set.

This study is the first to confirm that SIRI is an independent predictive biomarker for PHN development. The SIRI-based XGBoost model demonstrates excellent predictive performance, and the developed simplified risk scoring table has high clinical practicability, in which, early active intervention is recommended for patients with a risk score ≥18 to reduce the incidence of PHN.