Modelling and developing diabetic retinopathy risk scores on Indian type 2 diabetes patients

The objective was to develop diabetic retinopathy (DR) risk scores and compute prevalence and incidence probabilities of DR in Indian type 2 diabetes mellitus patients. A double sample of size 388 was collected from the R.G. Centre for Diabetes and Endocrinology, J.N.M.C., A.M.U., Aligarh, India, randomly distributed among training and test sets. DR risk scores of Iran and China were administered on Indian training set. Since prevalence probabilities of DR calculated by Logit model were unacceptable, thus actual data of Iranian and Chinese studies were simulated from their variable characteristics. Ridge regression was selected as optimal by regularization and cross-validation techniques. The yearly incidences of DR from ridge probabilities were determined using absorbing Markov chain. Receiver operating characteristic (ROC) curve and Hosmer Lemeshow test were exerted for model discrimination and calibration. Furthermore, these outcomes were implemented on the test sample. Out of 284 training sample patients, 23 had DR currently. Iranian score with an area of 0.815 (95% CI 0.765–0.859) was the better fit. Ridge coefficients acquired from Chinese simulated data contented the Indian data, providing accurate probabilities and an area of 0.784 (95% CI 0.731–0.830). Validating on test data, ROC curves for current, 1 year and 2 years prediction resulted in areas of 0.819, 0.811 and 0.686. Iranian score and simulated Chinese ridge coefficients for prevalence of DR were the best fit on Indian type 2 diabetes patients. Markov two-state model can be applied to forecast yearly incidence of DR.