Predicting diabetic retinopathy based on biomarkers: Classification and regression tree models

Abstract

Objective

We choose the classification and regression tree (CART) model as the analysis tool to identify clinical indicators that assess the risk factors for diabetic retinopathy (DR) and to determine the key risk factors associated with DR.

Methods

In this study, a total of 781 patients with diabetic mellitus(DM) were enrolled, including 395 patients with diabetic retinopathy(DR) and 386 patients without diabetic retinopathy. Univariate and multivariate logistic regression analysis were used to assess the risk factors of DR. Moreover, a machine learning approach, CART models were used to identify the high-order interactive effect of biomarkers and predict the DR.

Results

Out of 96 clinical test indicators, 11 were ultimately identified as the most critical. It was revealed in CART model analysis based on these 11 indicators that when urine creatinine(Ucr）≥ 127.75 mg/dl and cortisol ≥ 318.8 µg/l happened in DM patients, the prevalence of DR was 100 %, Even when UCr < 127.75 mg/dl, if cortisol ≥ 226.75 µg/l, 79.6 % of DM patients still have DR. Furthermore, even if neither of these two indicators exceeded the normal range, it increased the likelihood of DR when C-reactive protein (CRP) ≥ 10.11 mg/l. It was confirmed in the subgroup analysis with the CART model that there was an increase in CRP, a decrease in The 30-minute C-peptide level (C-P1), along with elevated cortisol, urinary protein (PRO), and UCr levels in identifying risk factors for DR. It was suggested the presence of kidney function impairment and insulin resistance, which triggered an inflammatory response, and ultimately exacerbated DR development increasing DR prevalence.

Conclusion

Based on the CART model analysis, routine serum biochemistry markers such as UCr, cortisol, and CRP, along with their respective thresholds, have been identified as potentially useful for identifying risk factors associated with DR at different the prevalence of DR rates. These findings might provided clinicians with a useful reference for the preliminary assessment of DR severity.