Development and validation of prediction models for diabetic retinopathy in type 2 diabetes patients

Le résumé fourni par la source

BACKGROUND AND OBJECTIVE: Prediction models enable healthcare providers to perform early risk stratification. This study aimed to develop and internally validate prediction models for 5- and 10-year risks of developing diabetic retinopathy (DR) in the Iranian individuals with type 2 diabetes. METHODS: This study utilized data from individuals with diabetes involved in the Shahroud Eye Cohort Study (ShECS), a prospective cohort study in Iran. The initial phase of ShECS began in 2009, with the second and third follow-up phases occurring in 2014 and 2019, respectively. Logistic regression developed prediction models, with bootstrap validation assessing internal validity. Model performance was evaluated using the discrimination and calibration. RESULTS: A total of 637 individuals with diabetes (35.0% men, mean (SD) of age: 53.0 (6.3 years)) were diagnosed. The five-year cumulative incidence of DR was 25.3% (95%CI: 21.8, 29.0%), and 17.0% (95%CI: 13.3, 21.0%) based on the second and third phases, respectively, while 10-year cumulative incidence was 40.0% (95%CI: 35.8, 44.0%). Incorporating various predictors, six models were developed with three recommended prediction models. Using mean blood pressure (MBP), non-fasting blood glucose (BG), and diabetes duration, Model-1 predicts 5-year risk indicating good calibration and discrimination with a c-statistic of 0.773 after bootstrap validation. The optimal statistical threshold was a predicted probability of 0.24. Model-2 predicts a 10-year risk incorporating diabetes duration, MBP, and BG, with a good calibration and a c-statistic of 0.687 after bootstrap validation showing moderate discrimination. The optimal statistical threshold was a predicted probability of 0.32. Model-3 predicts the 5-year risk using diabetes duration, MBP, glycated hemoglobin, high-density lipoprotein, triglycerides, and fasting blood glucose, showing good calibration and a c-statistic of 0.735 after bootstrap validation, indicating good discrimination. The optimal statistical threshold was a predicted probability of 0.20. CONCLUSION: Three prediction models with satisfactory performance were obtained using readily available predictors.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.