The research, led by Xia Sun and Tianrong Pan of the Second Affiliated Hospital of Anhui Medical University along with colleagues at Lishui Hospital of Traditional Chinese Medicine, took a retrospective look at 150 patients with type 2 diabetes who were treated at a tertiary hospital between January 2022 and December 2023. All patients met the 1999 World Health Organization diagnostic criteria for diabetes. Of the 150 individuals in the cohort, 92 had already developed diabetic retinopathy, giving the researchers a substantial group of both affected and unaffected patients to compare. The goal was not simply to catalog differences between the two groups but to build a working mathematical model that could estimate an individual patient’s risk.

The scientific rationale behind the study rests on a growing body of evidence that diabetic retinopathy is not merely a disease of high blood sugar damaging tiny retinal vessels. Subclinical inflammation and endothelial dysfunction—the gradual breakdown of the delicate inner lining of blood vessels—appear to play central roles in the earliest stages of retinal injury. Two families of biological markers have attracted attention in this context. The first is adipokines, signaling molecules secreted by adipose tissue that influence metabolism, inflammation, and vascular health. The second is hematological indices, particularly red cell distribution width, or RDW, a measure of the variation in the size of circulating red blood cells that has increasingly been recognized as a marker of systemic inflammation and oxidative stress.

Among the adipokines examined were omentin-1, adiponectin, apelin, and nesfatin-1, each with distinct and sometimes counterintuitive relationships to retinal disease. Using multivariate logistic regression, the researchers found that higher levels of omentin-1 and apelin were associated with an increased risk of diabetic retinopathy, while higher nesfatin-1 levels were associated with a reduced risk. Adiponectin, one of the most extensively studied fat-derived hormones, was incorporated into the final prediction model as well. The findings underscore how complex the biology of adipose tissue is in diabetes: molecules that are often described as protective in metabolic disease can behave differently when viewed through the lens of microvascular complications.

The hematological and biochemical picture that emerged was equally instructive. Patients with a mean arterial pressure above 120 mmHg faced elevated risk, as did those with abnormal levels of the liver enzymes aspartate aminotransferase and gamma-glutamyl transpeptidase, and those with high blood urea nitrogen, a marker of kidney function. Conversely, several factors appeared protective in the statistical analysis: higher high-density lipoprotein cholesterol, abnormal alanine aminotransferase, and high insulin resistance as measured by the homeostasis model assessment were each associated with reduced risk of retinopathy. Some of these associations—particularly the counterintuitive direction of the insulin resistance finding—highlight the hazards of interpreting retrospective data and the need for caution before drawing biological conclusions.

To translate these associations into a practical tool, the team constructed a prediction model and evaluated its performance using receiver operating characteristic analysis, the standard method for assessing how well a diagnostic or prognostic model separates those who develop a condition from those who do not. The apparent area under the curve, or AUC, was 0.921 with a 95 percent confidence interval of 0.880 to 0.962, a figure that in clinical terms suggests excellent discrimination. Because apparent performance almost always overstates how a model will behave in new patients, the researchers applied bootstrap internal validation, a resampling technique that estimates the degree of overfitting. The optimism-corrected AUC came out at 0.873 with a confidence interval of 0.821 to 0.925, still comfortably in the range considered good for a clinical prediction model.

Discrimination, however, is only half the story. A prediction model must also be well calibrated, meaning that when it says a patient has a 40 percent risk, roughly 40 percent of similar patients should actually develop the disease. The Hosmer-Lemeshow test suggested adequate calibration with a p value of 0.293, but the more visually informative calibration plot told a subtler tale. The model systematically underestimated risk in the intermediate probability range between 0.2 and 0.6 and overestimated risk around the 0.7 mark. The authors were candid about this discrepancy, describing the calibration as suboptimal and flagging it as an issue that would need to be addressed before any clinical deployment. This kind of transparency is increasingly valued in the prediction-model literature, where inflated claims of readiness have historically undermined trust in such tools.

The researchers also performed decision-curve analysis, a technique that quantifies the net benefit of using a model to guide clinical decisions across a range of risk thresholds. The model demonstrated positive net benefit across threshold probabilities of 10 percent to 60 percent, suggesting that within that band, acting on the model’s predictions would yield more benefit than the default strategies of treating all patients or treating none. In practical terms, this means an ophthalmology or endocrinology service could, in principle, use the model to prioritize patients for retinal screening, reserving scarce imaging and specialist resources for those whose calculated risk justifies closer surveillance.

Despite the encouraging numbers, the authors are careful to frame the study as a starting point rather than a finished clinical instrument. The cohort was modest in size, drawn from a single tertiary hospital, and analyzed retrospectively, all of which limit generalizability. The model has not yet been tested in an independent external cohort, the gold standard for demonstrating that a prediction tool performs as advertised outside the data that shaped it. The authors explicitly state that external validation and model recalibration are needed before clinical application. The calibration deviations observed in the intermediate risk range, where most clinical decisions are actually made, reinforce that recommendation.

Even with those caveats, the study adds to a compelling trend in diabetes research: the search for affordable, widely available biomarkers that can stratify risk before irreversible damage occurs. Adipokines and RDW can be measured from standard blood samples in most hospital laboratories, and combining them with blood pressure, liver enzymes, and kidney markers requires no new technology. If future prospective studies confirm the model’s performance and correct its calibration, the approach could help shift diabetic retinopathy care from reactive screening toward proactive risk management, catching the disease in time to preserve vision for the millions of people living with type 2 diabetes worldwide.

Subject of Research: A clinical prediction model for diabetic retinopathy in type 2 diabetes using adipokines and red cell distribution width

Article Title: A prediction model for diabetic retinopathy incorporating adipokines and red cell distribution width among patients with type 2 diabetes

Article References: Sun, X., Qiu, W., Chen, X., Zheng, P., Dan, L., Wu, J., & Pan, T. (2026). A prediction model for diabetic retinopathy incorporating adipokines and red cell distribution width among patients with type 2 diabetes. BMC Endocrine Disorders. https://doi.org/10.1186/s12902-026-02546-x

Image Credits: AI Generated

DOI: 10.1186/s12902-026-02546-x

Keywords: diabetic retinopathy, type 2 diabetes, adipokines, omentin-1, adiponectin, red cell distribution width, prediction model, apelin, nesfatin-1, logistic regression, ROC analysis, endothelial dysfunction