Establishment Of Prediction Model Of Insulin Medication In T2DM Patients Based On Data Mining

Objective: The number of diabetic patients in China is large,the harm is great,and the economic burden is heavy.Type 2 Diabetes Mellitus(T2DM)is the main type.However,individualized treatment has not been fully considered in the formulation of the initial plan of insulin at present.By collecting the medical data of T2 DM patients who need to use insulin drugs for a long time in the real world,and adopting the methods of data mining and machine learning,this study established the prediction model of T2 DM patients’ insulin treatment plan,in order to assist primary medical institutions in insulin treatment of T2 DM patients.Methods: This study collected medical data of T2 DM inpatients from the Department of Endocrinology,Geriatric Endocrinology and Caotang Endocrinology of Sichuan Provincial People’s Hospital from January 2016 to June 2019.After data preprocessing,a data set can be used for modeling.Then,the random forest algorithm was used for feature selection,and four machine learning algorithms,namely support vector machine,random forest,gradient boosting decision tree,and XGBoost,were used to establish the insulin medication plan,including species prediction model,dose prediction model,and proportion prediction model.The model with the best performance is selected through model evaluation and comparison.Results: After screening by inclusion and exclusion criteria,1048 target case data and 111 variables were obtained,including 31 qualitative variables and 80 quantitative variables.XGBoost showed the best performance in the species prediction model.The area under the micro average ROC curve was 0.74 and the area under the micro average ROC curve was 0.73.The XGBoost performed best in the dose prediction model.Its mean square error and average absolute error were 0.0175 and 0.0921,respectively.The absolute value of the difference between the 90% predicted value and its corresponding real value is less than 0.2014(u/kg),and the 95% is less than 0.2338(u/kg).XGBoost is still the best one in the proportion prediction model.Its mean square error and mean absolute error are 0.0420 and 0.1557,respectively.The absolute value of the difference between 90% prediction value and its corresponding real value is less than 0.2246,and the 95% is less than 0.3040.Conclusion: This study provides a new idea for the formulation of insulin medication plan,and the three models have a certain test value.After the model is further improved by adding data from other medical institutions,it can be widely used to realize its clinical application value.