Answers are grounded in this paper and cite their sources.
BACKGROUND: Pediatric chronic rhinosinusitis (CRS) frequently recurs after endoscopic sinus surgery (ESS), but reliable preoperative predictors are lacking, and adult-derived models may not apply to children. We developed and externally validated machine learning (ML) models to predict postoperative recurrence in pediatric CRS and identified key prognostic factors. METHODS: In this retrospective multicenter study, ML models-logistic regression (LR), random forest, XGBoost, and a stacking ensemble-were trained with class weighting on a development cohort of 633 pediatric ESS patients (aged 0-18 years), and an L1-optimized LR model was developed for variable selection from 33 candidate predictors. Performance was estimated by fivefold cross-validation, and the model was externally validated in 152 independent patients. Calibration was assessed using the Brier score, intercept, slope, and Hosmer-Lemeshow test, with logistic recalibration. RESULTS: Recurrence occurred in 242 of 633 patients (38.2%) in the development cohort and 46 of 152 patients (30.3%) in the external validation cohort. The L1-optimized LR model retained three predictors-nasal polyps, a higher Lund-Mackay CT score, and diffuse extent-each associated with higher recurrence risk, and achieved an AUC of 0.689, comparable to that of the other four class-weighted models. Laboratory findings and clinical symptoms were less predictive. On external validation, the model maintained discrimination (AUC = 0.756) but showed systematic miscalibration. Logistic recalibration restored calibration with essentially unchanged discrimination. CONCLUSIONS: The three-variable L1-optimized LR model provides a concise framework for preoperative recurrence risk stratification in pediatric CRS that could inform postoperative surveillance. A site-specific recalibration step is required before application to populations with different recurrence rates.
No takes yet. Share an insight, caveat, or question.
Lee et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: