目的：调查东南沿海地区翼状胬肉的患病率，分析相关危险因素并建立多模态机器学习预测模型。

方法：横断面流行病学调查。对舟山群岛居民群体随机抽样，通过现场眼部检查及流行病学调查，用单因素及多因素分析翼状胬肉相关的危险因素，并采用人工智能机器学习多模态模型对结构化与非结构化数据进行综合分析建立预测模型。

结果：本研究共调查1 842人，其中男926人(50.27%)，女916人(49.73%)； 平均年龄48.62±10.35岁。患翼状胬肉共644例(34.96%)，男323例(50.15%)，女321例(49.84%)； 其中单眼翼状胬肉者502例(77.95%)，双眼翼状胬肉者142例(22.05%)。无翼状胬肉者1 198例(65.04%)，其中男603例(50.33%)，女595例(49.67%)。职业、户外日照时长、睑板腺开口堵塞、年龄、吸烟、饮酒是引起翼状胬肉的危险因素，其中职业(OR=6.125，P<0.001)与户外日照时长(OR=5.348，P<0.001)是最危险因素； 其次是年龄(>40岁)(OR=5.295，P<0.001)及睑板腺开口堵塞(OR=5.248，P<0.001)。多模态机器学习模型预测效能良好，ROC曲线、PR 曲线、混淆矩阵均具备较高预测价值，其中AUC=0.86，AP=0.77。

结论：40岁以上，日照时长>4 h的渔民、农民是东南沿海地区患翼状胬肉的高发人群，睑板腺功能障碍与翼状胬肉的角膜侵入相关，可能促进其发展，应积极宣教与干预。多模态机器学习模型对翼状胬肉患病风险具有精准可靠的预测效能，有助于东南沿海人群翼状胬肉的风险分层与早期预警。

METHODS:A cross-sectional epidemiological study. A random sampling of the resident population of the Zhoushan Islands was performed. Data were collected through on-site ophthalmic examinations and epidemiological surveys. Univariate and multivariate analyses were performed to identify risk factors associated with pterygium. A multimodal artificial intelligence machine learning model was constructed to integrate and analyze both structured and unstructured data for predictive modeling.

RESULTS:A total of 1 842 participants were enrolled in this study, including 926 males(50.27%)and 916 females(49.73%), with a mean age of 48.62±10.35 y. A total of 644 patients(34.96%), including 323 males(50.15%)and 321 females(49.84%)were diagnosed with pterygium, including 502 cases(77.95%)were unilateral, and 142 cases(22.05%)were bilateral. There were 1 198 cases without pterygium, including 603 males(50.33%)and 595 females(49.67%). Occupation, daily outdoor sunlight exposure duration, meibomian gland orifice obstruction, age, smoking, and alcohol consumption were identified as significant risk factors for pterygium. Occupation and daily outdoor sunlight exposure duration were the most potent risk factors(OR=6.125, P<0.001; OR=5.348, P<0.001, respectively), followed by age >40 y(OR=5.295, P<0.001)and meibomian gland orifice obstruction(OR=5.248, P<0.001). The multimodal machine learning model demonstrated good predictive performance, as evidenced by the receiver operating characteristic(ROC)curve, precision-recall(PR)curve, and confusion matrix, all indicating high predictive value, AUC=0.86, AP=0.77.

CONCLUSION: Fishermen and farmers aged over 40 y with more than 4 h of daily sunlight exposure are at high risk for pterygium in the southeast coastal area. Meibomian gland dysfunction is correlated with corneal invasion of pterygium and may facilitate its progression; therefore, active health education and targeted intervention are recommended. The multimodal machine learning model provides accurate and reliable predictive capability for pterygium risk, which may facilitate risk stratification and early warning for pterygium in southeast coastal populations.