Predicting sublingual immunotherapy efficacy in allergic rhinitis

BACKGROUND: Sublingual immunotherapy (SLIT) efficacy for allergic rhinitis (AR) varies considerably, with 30%-40% of patients showing poor response. A reliable tool integrating multidimensional factors for individualized efficacy prediction remains lacking. This study aimed to construct an optimal prediction model incorporating clinical characteristics, environmental exposure factors, and immune-inflammatory indicators to predict SLIT efficacy in AR patients, and further establish a nomogram as an auxiliary interpretable tool for intuitive clinical application. MATERIALS AND METHODS: A total of 346 AR patients receiving SLIT were included and randomly allocated to training (n = 242) and validation (n = 104) cohorts at a 7:3 ratio. Baseline data included demographics, clinical features, symptom scores, environmental exposures, and immune-inflammatory indicators. Univariate and multivariate logistic regression analyses were performed to screen independent predictive factors. Three models, including random forest, support vector machine, and conventional logistic regression, were developed for performance comparison. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), calibration curves, and decision curve analysis (DCA). On the basis of independent predictors, a nomogram was constructed for visual interpretation. Shapley Additive Explanations (SHAP) analysis was further applied to interpret feature importance of the optimal model. RESULTS: Multivariate logistic regression confirmed these same seven variables as independent predictors of SLIT clinical efficacy in AR: disease duration, baseline symptom score, baseline medication score, air conditioning usage time, specific immunoglobulin E/total immunoglobulin E (sIgE/tIgE) ratio, interleukin (IL)-4, and IL-10 (all P < 0.05). The random forest model yielded optimal predictive performance, while the nomogram built on these predictors achieved acceptable discrimination with AUCs of 0.757 (95% CI: 0.676-0.837) in the training cohort and 0.729 (95% CI: 0.596-0.862) in the validation cohort, and showed better efficacy than the support vector machine (0.739) and conventional logistic regression (0.709) models. Calibration curves demonstrated good agreement between predicted probabilities and observed risks for the nomogram. DCA indicated that the model provided a high clinical net benefit across a wide range of threshold probabilities. SHAP analysis identified disease duration, sIgE/tIgE ratio, and baseline medication score as the three most influential features contributing to model predictions. CONCLUSION: The developed random forest model presents good discrimination, calibration, and clinical utility for predicting SLIT efficacy. The corresponding nomogram further enables intuitive individualized clinical assessment, providing a quantitative basis for personalized SLIT decision-making in AR patients.