Development and Validation of an Explainable Prediction Model for Postoperative Recurrence in Pediatric Chronic Rhinosinusitis.

Abstract

This study aims to develop an interpretable machine learning (ML) predictive model to assess its efficacy in predicting postoperative recurrence in pediatric chronic rhinosinusitis (CRS). A decision analysis was performed with retrospective clinical data. Recurrent group and nonrecurrent group. This retrospective study included 148 pediatric CRS treated with functional endoscopic sinus surgery from January 2015 to January 2022. We collected demographic characteristics and peripheral blood inflammatory indices, and calculated inflammation indices. Models were trained with 3 ML algorithms and compared their predictive performance using the area under the receiver operating characteristic (AUC) curve. Shapley Additive Explanations and Ceteris Paribus profiles were used for model interpretation. The final model was transformed into a web for interactive visualization. Among the 3 ML models, the Random Forest (RF) model demonstrated the best discriminative ability (AUC = 0.728). After reducing features based on importance and tuning parameters, the final RF model, including 4 features (systemic immune inflammation index (SII), pan-immune-inflammation value (PIV) and percentage of eosinophils (E%) and lymphocytes (L%)), showed good predictive performance in internal validation (AUC = 0.779). Global interpretation of the model suggested that L% and E% substantially contribute to the overall model. Local interpretation revealed a nonlinear relationship between the included features and model predictions. To enhance its clinical utility, the model was converted into a web (https://juice153.shinyapps.io/CRSRecurrencePrediction/). Our ML model demonstrated promising accuracy in predicting postoperative recurrence in pediatric CRS, revealing a complex nonlinear relationship between postoperative recurrence and the features SII, PIV, L%, and E%.