Exploratory development of prediction models for pharmacotherapy outcomes in trigeminal neuralgia: a combined analysis based on multi-source data.

BACKGROUND: Drug-refractory trigeminal neuralgia (DRTN) represents a formidable challenge in clinical management, with approximately 30%-50% of patients eventually progressing to DRTN. Early identification of high-risk DRTN populations and prediction of the timing of pharmacotherapy failure are crucial for optimizing clinical treatment strategies. METHODS: This retrospective study enrolled 163 primary trigeminal neuralgia patients receiving carbamazepine/oxcarbazepine with comprehensive imaging. A partially-blinded approach was used for data extraction: 135 patients (December 2022-December 2024) formed the primary cohort and 28 patients (December 2021-December 2022) served as the temporally separated validation cohort. Data comprised clinical baseline characteristics, pain and emotional scale scores, and imaging parameters. A nested cross-validation framework (20 random seeds) was used for model development, with feature selection confined to the training cohort within each seed. Five machine learning models were compared via pooled ROC curves aggregated across seeds, and SHAP analysis interpreted the optimal model at the median-performing seed. Age-unadjusted and age-adjusted multivariate Cox regression models were constructed, with four parallel univariate screening thresholds evaluated in sensitivity analyses. Model performance was assessed by Kaplan-Meier analysis, time-dependent ROC curves, and the concordance index. RESULTS: The support vector machine with radial basis function kernel (SVM-RBF) achieved the optimal performance, with average AUCs of 0.927 ± 0.041, 0.824 ± 0.049, and 0.806 ± 0.058 in the training, test, and validation cohort, respectively. In a sensitivity analysis using four univariate Cox screening thresholds, with age at first pain onset forced into all multivariate models, pain involved extent and medial temporal lobe atrophy (MTA) score were both identified as significant, independent candidate prognostic markers. Both Cox models showed moderate discriminative capacity for short-to-medium-term DRTN prognosis but attenuated long-term predictive efficacy. CONCLUSION: The SVM-RBF model, integrating multi-source heterogeneous data, demonstrated promising predictive performance for DRTN onset risk across the present cohort. The Cox models showed moderate discriminative capacity for short-to-medium-term prognosis, though long-term predictive accuracy was attenuated. These findings provide preliminary evidence for early risk stratification and individualized treatment of DRTN, while underscoring the need for improved survival prediction strategies to address the limited long-term performance of the current Cox models.