Predicting Surgical Outcomes in Chronic Rhinosinusitis From Preoperative Patient Data: A Machine Learning Approach

Endoscopic sinus surgery (ESS) fails to adequately address symptoms in some chronic rhinosinusitis (CRS) patients. This study aims to evaluate machine learning models (MLMs) in predicting ESS outcomes from preoperative data.

CRS patients who underwent primary ESS were used to train and test MLMs for the identification of the minimal clinically important difference (MCID) between pre- and post-operative total SNOT-22 scores. Fifty-nine preoperative predictors were included in the MLMs. Each model was optimized using K-fold cross validation. The performance of two MLMs and a conventional logistic regression (LR) model was evaluated.

The study cohort constituted 242 patients. The Ensemble model achieved the highest discriminative performance (Ensemble area-under-the-curve [AUC] 0.89, 95% CI 0.87-0.91; LR AUC 0.85, 95% CI 0.83-0.87; XGBoost AUC 0.74, 95% CI 0.72-0.76, p < 0.05) and accuracy (Ensemble 87.8%; LR 81.6%; XGBoost 81.6%). The XGBoost model achieved the highest sensitivity (XGBoost 97.1%; LR 76.5%; Ensemble 85.3%) but poorer specificity (XGBoost 46.7%; LR 93.3%; Ensemble 93.3%). The most predictive factors in the Ensemble model were age, preoperative SNOT-22 and Patient Health Questionnaire 2 (PHQ-2) scores, nasal obstruction, and facial pain/pressure.

The Ensemble MLM achieved the highest overall accuracy and discriminative performance and matched LR for the highest specificity, suggesting that this model could be used in clinical decision-making to "rule in" patients likely to achieve MCID after ESS. This is the highest performance achieved in an MLM trained toward clinical outcomes in CRS. After validation in a multicenter cohort, the Ensemble MLM could become a useful adjunct in clinical decision-making for ESS.