Abstract: Background：Retinopathy of prematurity (ROP) can cause irreversible visual impairment in severe cases, and there is a lack of appropriate risk prediction model for identifying severe ROP specifically tailored to the Chinese preterm population. Objective：To construct and validate an early risk prediction model for severe ROP in premature infants with gestational age ≤ 34 weeks based on the Chinese neonatal population. Design：Retrospective cohort study. Methods：From January 2020 to March 2022, premature infants admitted to the Neonatal Intensive Care Unit (NICU) with a gestational age of ≤ 34 weeks and who underwent a complete eye examination were included. The cohort was divided into training and validation sets using the June 2021 admission time as the cutoff point. Clinical information was collected at the time of the first postnatal eye examination, and regular screenings and followups were conducted according to the ROP screening guidelines. Severe ROP is defined as having stage 3 or higher lesions, positive for plus disease, or requiring laser and cryotherapy treatment. All other cases are categorized as mild ROP. In the training set, clinical variables with collinearity were excluded based on the Pearson correlation coefficient and the final predictive factors were determined through Lasso regression and logistic regression analysis, then the prediction model was constructed and presented as a nomogram. In the validation set, the area under the curve (AUC) of receiver operating characteristic (ROC) was used to evaluate the discrimination and determine the optimal cutoff value of the prediction model, in which the sensitivity and specificity of the model were calculated. Based on the total population, clinical efficacy of the model was finally assessed by decision curve analysis (DCA). Main outcome measures：The predictive performance of the model for identifying premature infants who will progress to severe ROP. Results：There were 1,380 cases in the training set (76 cases with severe ROP) and 592 cases in the validation set (36 cases with severe ROP). The differences in the distribution of baseline information between these two groups were not statistically significant. The nomogram prediction model established by including five predictors of gestational age, birth weight, prenatal corticosteroids, invasive mechanical ventilation, and lateonset sepsis had an AUC of 0.87 (95% CI: 0.83 to 0.90) and 0.88 (95% CI: 0.82 to 0.94) in the training and validation sets, respectively. When the predictive threshold was set at 0.04, the model had a sensitivity of 0.833 (95% CI: 0.7120.955) and a specificity of 0.786 (95% CI: 0.7520.820) in the validation set. The clinical decision curve demonstrates that the model has good clinical utility and provides higher net benefits when the threshold probability ranges from 5% to 40%. Conclusions：The severe ROP prediction model established has a certain reference value for evaluating the risk of severe ROP in neonates with gestational age ≤ 34 weeks.

Key words: Retinopathy of prematurity, Lasso regression, Prediction model, Nomogram