# Risk prediction models for diabetic retinopathy: a systematic review

## TL;DR

This paper reviews predictive models for diabetic retinopathy to help healthcare professionals identify high-risk patients and improve model development.

## Contribution

A systematic review of 15 predictive models for diabetic retinopathy, assessing their performance and risk of bias.

## Key findings

- The 15 models showed area under the curve values ranging from 0.700 to 0.960.
- All models had high risk of bias, but five had better applicability.
- Common risk factors included diabetes duration, age, and glycosylated hemoglobin.

## Abstract

Diabetic retinopathy, a prevalent complication of diabetes mellitus, is a growing public health concern. The use of robust predictive models can aid healthcare professionals in identifying high-risk patients, enabling them to implement early intervention and treatment strategies.

To systematically evaluate published prediction models for diabetic retinopathy, select better prediction models for healthcare professionals, and provide a valuable reference for model optimization.

A comprehensive search was conducted across the PubMed, Web of Science, Embase, and the Cochrane Library databases for relevant literature on predictive models for diabetic retinopathy. The search period was set from the time of library construction to November 14, 2023. Furthermore, risk of bias and applicability assessment of the included study models were performed using the PROBAST risk assessment tool.

A total of 2030 studies were retrieved, including 15 studies. The range of the working characteristic curve of the subjects for the 15 models varied from 0.700 to 0.960. All 15 included studies were recognized as high risk of bias. However, five studies had better applicability. The 15 models had Common risk factors for the 15 models included diabetes duration, age, glycosylated hemoglobin, serum creatinine and urinary albumin creatinine ratio.

While the performance of the 15 models had certain predictive performance, the high risk of bias is a concern. Hopefully, future studies will ensure transparency and science in the model-building process by conducting large-sample integrated machine learning, reinforcing multicenter external validation.

This study was registered with PROSPERO, an international prospective systematic evaluation registry platform, and the title was approved with registration number CRD42023483749.

https://www.crd.york.ac.uk/PROSPERO/, identifier CRD42024559392.