Abstract:Objective： To construct a reliable risk prediction model for postoperative dry eye disease (DED) in glaucoma patients using the XGBoost algorithm. Methods：A retrospective analysis was performed on 300 glaucoma patients who received surgical treatment at our hospital from July 2022 to June 2025. All patients were divided into a DED group and a non-DED group according to the occurrence of postoperative dry eye disease. Clinical data of the patients were collected, and univariate and multivariate logistic regression analyses were employed to screen out the independent risk factors for postoperative DED. The patients were randomly allocated into a training set (210 cases) and a validation set (90 cases) at a ratio of 7:3. An XGBoost risk prediction model was built with the identified independent risk factors as feature variables, and the SHapley Additive exPlanations (SHAP) bar plot and beeswarm plot were used for visual interpretation of the model. The predictive efficacy of the model was evaluated via receiver operating characteristic (ROC) curve analysis. Results： Among the 300 glaucoma patients (300 eyes), 96 developed postoperative DED, resulting in an incidence rate of 32.00%. The non-DED group comprised 204 patients (204 eyes) (104 males and 100 females), and the DED group comprised 96 patients (96 eyes) (55 males and 41 females). Univariate analysis revealed statistically significant differences between the DED and non-DED groups in terms of age, comorbid diabetes, meibum viscosity grade, tear film breakup time, meibomian gland dysfunction, and operative time (all P＜0.05). Multivariate logistic regression analysis confirmed that all the above factors were independent risk factors for postoperative DED (all P＜0.05). The XGBoost model showed that these independent risk factors were ranked in descending order of predictive importance as: tear film breakup time, comorbid diabetes, operative time, age, meibomian gland dysfunction, and meibum viscosity grade. ROC curve analysis demonstrated that the area under the curve (AUC) of the XGBoost model was 0.84 (95%CI: 0.78-0.90) for the training set and 0.83 (95%CI: 0.74-0.92) for the validation set, with both values showing statistical significance (all P＜0.05). Conclusion： The XGBoost risk prediction model established in this study exhibits favorable predictive performance for postoperative DED in glaucoma patients. This model can be clinically applied to accurately identify high-risk patients with postoperative DED, and provide a reliable theoretical basis for clinicians to implement targeted preventive and interventional measures, thus optimizing the perioperative eye surface management of glaucoma patients.
