Establishment of Clinical Prediction Model Based on the Study of Risk Factors of Stroke in Patients With Type 2 Diabetes Mellitus

Purpose: Stroke has sparked global concern as it seriously threatens people's life, bringing about dramatic health burdens on patients, especially for type 2 diabetes mellitus (T2DM) patients. Therefore, a risk scoring model is urgently valuable for T2DM patients to predict the risk of stroke incidence and for positive health intervention. Methods: We randomly divided 4,335 T2DM patients into two groups, training set (n = 3,252) and validation set (n = 1,083), at the ratio of 3:1. Characteristic variables were then selected based on the data of training set through least absolute shrinkage and selection operator regression. Three models were established to verify predictive ability. Foundation model was composed of basic information and physical indicators. Biochemical model consisted of biochemical indexes. Integrated model combined the above two models. Data of three models were then put into logistic regression analysis to form nomogram prediction models. Tools including C index, calibration plot, and curve analysis were implemented to test discrimination, calibration, and clinical use. To select the best predicting model, net reclassification improvement (NRI) and integrated discrimination improvement (IDI) were put into effect. Results: Eleven risk factors were determined, including age, duration of T2DM, estimated glomerular filtration rate, systolic blood pressure, diastolic blood pressure, low-density lipoprotein, high-density lipoprotein, triglyceride, body mass index, uric acid, and glycosylated hemoglobin A(1c), all with significant P-values through logistic regression analysis. In the training set, areas under the curve of three models were 0.810, 0.819, and 0.884, whereas in the validation set, they were 0.836, 0.832, and 0.909. Through calibration plot, the S:P values in the training set were 0.836, 0.754, and 0.621 and were 0.918, 0.682, and 0.666 separately in the validation set. In terms of the decision curve analysis, the risk thresholds were, respectively, 8–73%, 8–98%, and 8%~ in the training set and 8–70%, 8–90%, and 8–95% in the validation set. With the aid of NRI and IDI, integrated model is proved to be the best model in training set and validation set. Besides, internal validation was conducted on all the subjects in this study, and the C index was 0.890 (0.873–0.907). Conclusion: This study established a model predicting risk of stroke for T2DM patients through a community-based survey.