Enhancing Long-Term Recurrence Prediction in Chronic Rhinosinusitis Following Endoscopic Sinus Surgery Using a Discrete-Time Model.

Abstract

To compare the prognostic performance of a discrete-time pooled logistic regression (PLR) model and a continuous-time Cox proportional hazards (Cox) model for predicting 2-, 5-, 10-, and 15-year recurrence risk after endoscopic sinus surgery (ESS) in chronic rhinosinusitis (CRS). Retrospective cohort study. Single tertiary academic center. A total of 543 CRS patients who underwent ESS were included for analysis of time to recurrence. Ten predictors, including age, sex, smoking status, asthma, NSAID hypersensitivity, symptom duration, blood eosinophil count (BEC), nasal polyp score (NPS), MLK Discharge-Edema subscore (MLK-DE subscore), and Lund-Mackay (LM) score, were evaluated using LASSO-penalized PLR and Cox models. Nomogram performance was assessed using time-dependent AUROC, Brier scores, calibration, 1000-bootstrap internal validation, and decision curve analysis (DCA). Recurrence occurred in 46.8%. LASSO identified age, NSAID hypersensitivity, asthma, symptom duration, BEC, NPS, MLK-DE subscore, and LM score as key predictors, with smoking status retained only in PLR model. PLR model demonstrated numerically higher discrimination (AUROCs 0.899-0.912 vs 0.879-0.899) with similar Brier scores (0.111-0.144 vs 0.128-0.142). PLR calibration remained strong over time, whereas Cox calibration declined over longer follow-up. Although Cox performed slightly better at early time points, PLR demonstrated consistent discrimination during internal validation (AUROCs 0.863-0.892 vs 0.876-0.897). DCA showed greater and stable net benefit for PLR across all time points. Both models performed well in predicting long-term recurrence, but the LASSO-penalized PLR nomogram provided numerically higher and more consistent discrimination, calibration, and clinical utility over time.