A Novel Machine Learning Signature Incorporating Lipid and Inflammatory Biomarkers to Predict Pulsed Radiofrequency Efficacy in Zoster-Associated Pain.

INTRODUCTION: Pulsed radiofrequency (PRF) is a pivotal neuromodulation strategy for zoster-associated pain (ZAP); however, clinical outcomes exhibit significant interindividual heterogeneity. This study aimed to identify robust predictors and develop a transparent machine learning (ML) framework to forecast treatment response, thereby facilitating individualized pain management. METHODS: We conducted a retrospective analysis of a large-scale cohort comprising 1773 patients with ZAP treated with PRF. Patients were stratified into "responders" and "nonresponders" on the basis of clinical outcomes at a 3-month follow-up. To handle an initial pool of 47 multidimensional clinical and laboratory variables, a tripartite feature selection pipeline-incorporating least absolute shrinkage and selection operator (LASSO), Boruta, and multivariable logistic regression-was implemented. We benchmarked eight ML architectures. The optimal model was interpreted using Shapley additive explanations (SHAP) to ensure biological transparency and subsequently deployed as an interactive point-of-care tool. RESULTS: The favorable clinical response rate for PRF was 68.0%. A parsimonious set of five core predictors was identified: age, baseline Numerical Rating Scale (NRS) score, preoperative opioid use, apolipoprotein B (ApoB), and neutrophil-to-monocyte ratio (NMR). Among the candidate algorithms, the CatBoost architecture was selected for its robust performance, achieving the highest F1 score (0.853) and an area under the receiver operating characteristic curve (AUROC) of 0.834. SHAP analysis revealed that preoperative opioid use was the most potent determinant of suboptimal response, followed by advanced age and high baseline pain intensity. Notably, elevated ApoB levels emerged as a novel metabolic indicator of favorable prognosis, whereas an increased NMR suggested a systemic pro-inflammatory state associated with diminished PRF efficacy. CONCLUSIONS: We developed and internally validated a high-performance CatBoost-based model for predicting PRF outcomes in ZAP. By integrating novel metabolic (ApoB) and immunoinflammatory (NMR) biomarkers with established clinical metrics, this model provides a granular approach to risk stratification. The deployment of a Streamlit-based calculator translates complex algorithmic insights into an accessible clinical decision-support system.