关键词: 近视, 眼底病变, 预测模型, 系统评价

Abstract:  Objective Systematic review of risk prediction models for fundus lesions in myopic patients.Design Systematic review. Participants Literature on risk prediction model of myopia-related fundus lesions. Methods A computerized search was conducted in CNKI, Wanfang Data, CBM, PubMed, Web of Science, Embase, CINAHL, and the Cochrane Library for studies related to prediction models of fundus lesions in myopic patients, covering the period from database inception to August 2025. Data were extracted from the CHARMS list, and the PROBAST tool was used to assess the risk of bias and suitability of the study. If any one of the four items, namely study subjects, predictors, outcomes, and statistical analysis, was rated as "high risk", then the model was considered to be at "high risk". If one or more of the three items, study subjects, predictors, and outcomes, are rated as "No" or "Possibly No", then the model was at high risk in terms of applicability. The area under the working characteristic curve (AUC) of the subjects was used to evaluate the model differentiation with reference to general empirical criteria. AUC≥0.70 was considered an acceptable degree of discrimination. Extract and record the validation type (internal/external validation) of the model, and summarize the most important predictors. Main Outcome Measures Model performance, predictors, validation, bias risk and applicability of the model were analyzed. Results A total of 10 studies involving 24 models were included. The models demonstrated acceptable discrimination (AUC: 0.73~0.95). However, only five studies underwent both internal and external validation, indicating insufficient independent validation. Based on the PROBAST tool evaluation, 7 studies were rated as high risk of bias due to single-center small sample size and incomplete measurement procedures for predictors, and 6 studies were rated as high applicability risk because they rely on specific techniques or equipment that were difficult to popularize in clinical practice. Key predictors included axial length, refractive error, age, retinal and choroidal thickness, ocular complications, and laboratory inflammatory markers. Conclusions Most of the existing 10 risk prediction models for myopia-related fundus lesions demonstrate good discriminative power, but over half exhibit high bias risks, and half lack external validation with suboptimal methodological quality. These models are not recommended for direct application in routine clinical decision-making.

Key words:  Myopia, Fundus lesions, Prediction model, Systematic review