Abstract
Blood glycosylated hemoglobin level can be affected by various factors in patients with type 2 diabetes and cardiovascular diseases. Frequent measurements are expensive, and a suitable estimation method could improve treatment outcomes. PATIENTS AND METHODS: 93 patients were recruited in this research. We analyzed a number of parameters such as age, glucose level, blood pressure, Body Mass Index, cholesterol level, echocardiography et al. Patients were prescribed metformin. One group (n=60) additionally was taking sitagliptin. We applied eight machine learning methods (k nearest neighbors, Random Forest, Support Vector Machine, Extra Trees, XGBoost, Linear Regression including Lasso, and ElasticNet) to predict exact values of glycosylated hemoglobin in two years. RESULTS: We applied a feature selection approach using step-by-step removal of them, Linear Regression on remaining features, and Pearson's correlation coefficient on the validation set. As a result, we got four different subsets for each group. We compared all eight Machine Learning methods using different hyperparameters on validation sets and chose the best models. We tested the best models on the external testing set and got R(2) = 0.88, C Index = 0.857, Accuracy = 0.846, and MAE (Mean Absolute Error) = 0.65 for the first group, R(2) = 0.86, C Index = 0.80, Accuracy = 0.75, and MAE = 0.41 for the second group. CONCLUSION: The resulting algorithms could be used to assist clinical decision-making on prescribing anti-diabetic medications in pursuit of achieving glycemic control.