Construction Of A Hypoglycemia Prediction Model Based On A Machine Learning Algorithm For Hospitalized Type 2 Diabetes Patients

Objective1.To initially determine the risk factors for the occurrence of hypoglycemia in hospitalized type 2 diabetic patients.2.To construct a hypoglycemic risk prediction model based on the risk factors of hypoglycemia in hospitalized type 2 diabetic patients obtained from Meta-analysis and to screen the best model from it,to provide a scientific and effective assessment tool for preventing the occurrence of hypoglycemia in hospitalized T2 DM patients,to provide a theoretical basis for medical personnel to screen high-risk patients and take targeted preventive and nursing measures promptly.Methods1.CNKI(China Knowledge Network),VIP(Vip),Wan Fang Data(Wan Fang),Pubmed,Embase,The Cochrane Library,and Web of Science databases were searched to screen the literature related to risk factors for hypoglycemia in hospitalized type 2diabetes mellitus,and data were extracted and cross-checked by two investigators,and The NOS scale was used to evaluate the quality of the literature,and Meta-analysis was performed by Rev Man 5.4 software to initially determine the factors influencing hypoglycemia.2.Based on the influencing factors obtained from Meta-analysis,the influencing factors of hypoglycemia to be collected were determined by combining the opinions of clinical experts in the endocrinology department,and a data collection form was developed.Retrospectively,all patients with type 2 diabetes admitted to the endocrinology department of a tertiary hospital in Anhui province from January 2021 to January 2022 were surveyed.The electronic medical records of the hospital case room were retrieved by the investigator himself through the patient’s hospitalization number,and the general data,disease-related data,and laboratory test data were obtained according to the nadir criteria.The patients’ general data,disease-related data,and laboratory test data were obtained according to the criteria,and the outcome variable hypoglycemia was obtained through the OA hypoglycemia reporting system;The data with more than 25% missing values were excluded,and the remaining data were filled according to the type of variables,and the variables with statistical differences(P<0.05)were screened by one-way analysis.3.The variables with statistical differences(P <0.05)in the univariate analysis were used as input variables for the model,and whether hypoglycemia occurred was used as the model label,and Logistic Regression,Random Forest Classifier,Decision Tree Classifier,and XGB were used respectively,After the models were constructed,the accuracy,precision,recall,F1 score,and AUC values were used to evaluate the prediction performance of the models,and then the optimal model was selected.Results1.Meta-analysis results: 14 papers were finally included after the screening,and the results showed that there were statistically significant differences between the two groups for a total of 8 indicators,suggesting that age,disease duration,glycosylated hemoglobin,body mass index,insulin use,gender,urinary microalbumin and hospitalization days were risk factors for hypoglycemia in hospitalized type 2 diabetic patients,and there was no statistically significant difference between the two groups for1 indicator,suggesting that blood creatinine level was not related to the occurrence of hypoglycemia in hospitalized diabetic patients.2.Model construction results: In this study,612 cases of study subjects were collected,including 115 cases in the hypoglycemic group and 497 cases in the nonhypoglycemic group,and the incidence of hypoglycemia was 18.79%.The age of the patients ranged from 18 to 88 years old,and there were 201 cases(32.8%)in women and 411 cases(67.2%)in men.Univariate analysis showed that 19 variables were statistically significant: blood potassium,hemoglobin,waist circumference,glycated serum protein,fasting insulin,alanine aminotransferase,aspartate aminotransferase,blood urea nitrogen,triacylglycerol,HDL cholesterol,total bilirubin,urinary creatinine,2h postprandial C-peptide,fasting C-peptide,disease duration,BMI,diabetic nephropathy,hyperlipidemia,and treatment regimen.The variables obtained from the univariate analysis were included in the logistic regression model,XGboost model,decision tree model,and random forest model and validated in the test set.The AUCs of the logistic regression model,XGboost model,decision tree model,and random forest model in the training set were 0.82,0.91,0.84,and 0.89,respectively;in the test set The AUCs of the logistic regression model,XGboost model,decision tree model,and random forest model in the test set were 0.76,0.80,0.73 and 0.78,respectively.combining the accuracy,precision,recall,and F1 score,it can be concluded that the prediction performance of the hypoglycemia prediction model constructed by XGboost algorithm is the best.After ranking the importance of features by the XGboost,eight features,fasting insulin,urinary creatinine,glycated serum protein,blood urea nitrogen,2h postprandial C-peptide,total bilirubin,fasting C-peptide,and HDL cholesterol,were the most important influencing factors.All statistical analyses were performed using Revman,SPSS,and Python software.ConclusionAmong the hypoglycemia prediction models constructed based on machine learning algorithms for inpatients with type 2 diabetes,the hypoglycemia prediction model constructed by the XGboost algorithm had the best predictive efficacy in comparison.Eight characteristics of fasting insulin,urinary creatinine,glycated serum protein,blood urea nitrogen,2h postprandial C-peptide,total bilirubin,fasting C-peptide,and high-density lipoprotein cholesterol were the most important influencing factors,which could provide some help for the early identification of hypoglycemia.