Construction and Validation of a Predictive Model for Voriconazole Treatment Efficacy in Influenza-Associated Pulmonary Aspergillosis Based on Machine Learning

Abstract

Purpose To analyze risk factors affecting voriconazole efficacy in influenza-associated pulmonary aspergillosis (IAPA) and build interpretable machine learning (ML) predictive models. Patients and Methods This retrospective study enrolled voriconazole-treated IAPA patients into a primary retrospective set (Jan 1, 2020–Jul 31, 2025) and a separate temporal external validation set (Aug 1, 2025–Apr 30, 2026). To avoid data leakage, the primary set was stratified by treatment response and randomly split at a 7:3 ratio into training and internal validation set prior to feature screening. Forty-two baseline candidate variables were collected. Feature selection was performed exclusively within the training dataset via univariate screening, 10-fold cross-validated Least Absolute Shrinkage and Selection Operator (LASSO), and Variance Inflation Factor (VIF) collinearity filtering. Parallel multivariate logistic regression was performed to identify independent predictors without additional stepwise elimination. Validation sets were only used for model performance evaluation. Seven ML algorithms were constructed, and SHapley Additive exPlanations (SHAP) analysis was applied to interpret the best-performing model. Results Patients in the internal set were categorized into voriconazole-responsive (VR, n=284) and non-responsive (VNR, n=56) groups based on clinical remission, radiological absorption, and reduction of inflammatory biomarkers. Multi-step variable screening retained 11 non-collinear candidate predictors. Multivariate logistic regression identified four independent predictors for poor therapeutic response: ICU admission, concomitant midazolam use, elevated baseline neutrophil percentage, and reduced baseline platelet count. The CatBoost model achieved the optimal discrimination, with internal-validation AUC = 0.88 (95% CI: 0.87–0.88) and temporal external-validation AUC = 0.83 (95% CI: 0.70–0.94). Decision Curve Analysis (DCA) demonstrated that all seven models yielded positive net clinical benefit across the threshold probability range of 0.00–0.90. SHAP analysis further illustrated the important contribution of ICU admission and midazolam to prediction outcomes, with partial discrepancy from multivariate logistic regression feature ranking. Conclusion This study identified four independent predictors of voriconazole efficacy in IAPA patients. The established CatBoost model delivers stable, interpretable prediction to guide individualized voriconazole therapy for IAPA.