OBJECTIVES : In the real-world environment, some patients with T2DM failing to receive timely treatment intensification, which will lead to poor glycaemic control and increase the risk of complications. However, there is little clinical research about these patients. If machine learning (ML) technology can be used to recognize and early manage high-risk patients, the control of glycaemic can be improved. Therefore, the aim of the study is to develop prediction models incorporating glycosylated hemoglobin (HbA1c) and complications for patients with type 2 diabetes (T2DM) under long-term treatment.

METHODS : We collected demographic, disease, and medication information, laboratory tests, and economic data from 165 patients with T2DM who had neither been monitored for HbA1c nor had treatment changes for at least one year. Set complications and HbA1c as the target variables. Seven types of ML algorithms were used to establish 18 prediction models. Ten independent replicate experimental data were obtained for each model. The whole data set was divided into a training set and a testing set in a ratio of 8:2. The predicted performance was evaluated by the area under the receiver operating characteristic (ROC) curve (AUC) of the testing set.

RESULTS : Among 165 T2DM patients under long-term hypoglycemic treatment, 83 had complications while 36 patients’ s HbA1c were under control (<7%). The best models for diabetic nephropathy, diabetic peripheral neuropathy, diabetic vascular disease, diabetic eye disease, and HbA1c were the ensemble model (AUC, 0.9024 ± 0.0405), the discriminant model (AUC, 0.8585±0.0502), the ensemble model (AUC, 0.8889 ± 0.0598), the discriminant model (AUC, 0.8319±0.0856) and the Bayesian model (AUC 0.8249±0.0918), respectively.

CONCLUSIONS : Our machine learning prediction models had acceptable values of sensitivity and specificity, and can potentially be used to predict the complications and HbA1c in patients with T2DM under long-term treatment.