[Development and validation of a machine learning-based prognostic model for chronic rhinosinusitis with nasal...

Objective: To develop predictive models based on multiple machine learning (ML) algorithms, to systematically compare the predictive performance of these models in assessing the risk of poor postoperative prognosis in patients with chronic rhinosinusitis with nasal polyps (CRSwNP), and to further explore their potential clinical application value. Methods: A total of 333 CRSwNP patients who underwent endoscopic sinus surgery were prospectively enrolled from the Department of Otorhinolaryngology and Head and Neck Surgery, Renmin Hospital of Wuhan University, from March 2021 to July 2024. All patients were randomly allocated into a training set (n=233) and a test set (n=100) at a 7∶3 ratio. Preoperative clinical data and 1-year follow-up data were collected. Medication status, clinical symptoms and endoscopic findings were assessed according to the European Position Paper on Rhinosinusitis and Nasal Polyps 2020 (EPOS 2020), and patients were classified into favorable and poor prognosis groups. Differential analysis was conducted to screen independent prognostic factors. A total of 113 machine learning models were established using 12 feature selection algorithms, with prognostic factors as feature variables for model training. Model performance was assessed using the area under the curve (AUC) derived from 10-fold cross-validation in the training set. Model reliability was validated in the test set, and SHAP values were used to rank the importance of predictive features. Results: At the 1-year follow-up of 333 CRSwNP patients, the incidence of poor postoperative prognosis was 36.34%. Compared with patients in the favorable prognosis group, patients in the poor prognosis group were older, with a higher incidence of olfactory cleft polyps and a higher rate of prior alcohol consumption. They also had elevated peripheral blood and tissue eosinophil counts, higher Lund-Mackay CT scores, Visual Analogue Scale (VAS) scores and olfactory scores, as well as elevated levels of interleukin (IL)-2, IL-6, IL-10, IL-17, tumor necrosis factor α (TNF-α) and interferon γ (IFN-γ). In contrast, peripheral blood and tissue neutrophil counts, serum immunoglobulin M (IgM) and complement C3 levels were markedly decreased (all P<0.05). Among the 113 machine learning models constructed using 12 algorithms, the ensemble model of generalized linear model boosting plus gradient boosting machine (glmBoost+GBM) yielded the optimal performance, with a mean AUC of 0.894 in the training set. Key predictive features included olfactory cleft polyps, advanced age, elevated local eosinophils, reduced tissue neutrophils, elevated baseline VAS scores, and baseline olfactory impairment. In an independent test set, the model achieved an AUC of 0.813 (95%CI: 0.728-0.899) with good calibration. Decision curve analysis further verified its favorable clinical application value. Conclusions: The glmBoost+GBM ensemble learning model combined with feature selection can effectively predict postoperative prognosis in CRSwNP patients, with good predictive efficacy and interpretability. This model provides a reliable tool for clinical risk stratification of CRSwNP patients.