Development and validation of sex-specific clinical prediction models for treatment-required retinopathy of prematurity.

Abstract

Purpose: To develop clinically practical, sex-specific prediction models for identifying TR-ROP, irrespective of fundus photography, and evaluate its generalizability, efficiency, productivity, and interpretability.

Methods: We selected premature infants who suffered risk of TR-ROP and received fundus examination between 2012 and 2022. Logistic Regression (LR) Model, Random forest-LR Model and LASSO-LR Model were constructed and the model with the best performance was chosen for predictions of the occurrence of TR-ROP.

Results: Among 7,235 preterm infants received ROP screening, 408 (5.63%) were TR-ROP. The median follow-up time was 24 months. Male and female shared some modifiable risk and protective factors, but they also presented independent risk factors. The sex-specific model based on birth weight, gestational age, hypoxic ischemic encephalopathy, multiple births, blood transfusion (male) and birth weight, gestational age, head circumference, cesarean delivery, blood transfusion (female) were selected by LR showed more promising results in the prediction of TR-ROP in the internal validation cohort (male: AUC 0.855-0.981, specificity 0.895; female: AUC 0.950-0.995, specificity 1.000). The performance of the above sex-specific models also demonstrated performance in the external validation cohorts (male: AUC 0.806-0.951, specificity 0.824; female: AUC 0.625-0.919, specificity 0.727). The C-index showed the sex-stratified models displayed better clinical predictive utility than the overall model.

Conclusion: Our study provides a sex-specific clinical risk prediction tool for TR-ROP, which may help preterm infants identify their potential risk profile, reduce unnecessary fundus examination and provide guidance to prevent disease progression.