Objective: To identify essential predictors to build a cost-effective prediction model for early detection of diabetic retinopathy (DR).
Methods: We performed a retrospective case/control cohort study based on data extracted from a large electronic health records (EHR) database. Univariate association was used to evaluate patient demographics and 26 lab results in relation to incidence of DR within a 6-month window in future. Backward variable selection was implemented with logistic regression to develop a model based on significantly associated variables. The model was further simplified by narrowing down to shared predictors with previous models in literature. Highly correlated lab variables were further investigated to evaluate the impact of substituting or removing them from the model.
Results: A case cohort of 3,817 DR patients and a control cohort consisted of an equal number of non-DR diabetic patients were studied. Age, gender, race, and 22 lab values were found to be significantly associated with DR. The full model based on the 25 variables had AUC=0.817. Age, gender, race, and 4 lab variables - HbA1c, creatinine, red blood cell (RBC) count, and white blood cell count were efficient to predict DR. The AUC of the reduced model based on the 7 variables was 0.8, which was close to the accuracy of the full model. Furthermore, RBC could be replaced by its highly correlated equivalents, hematocrit (correlation coefficient r=0.9) and hemoglobin (r=0.86), without noticeably impacting the accuracy (AUC with variable substitution ranged from 0.798 to 0.801).
Conclusion: HbA1c, creatinine, RBC, and WBC were found to be key lab variables that can form an accurate DR prediction model. RBC could be replaced by one of their correlated equivalents. The choice of which variables are most cost-effective and efficient to include will depend on the expense and informativeness of the lab test performed in various clinical settings. Future work of this study will focus on external validation of proposed model.
R. Wang: None. Z. Miao: None.