Research On Insulin Evaluation Models And Algorithm Based On Machine Learning

Insulin was a self secreted hormone that could decompose glucose.It was produced by human islets β cells were secreted by exogenous stimulation.Once the function of the cells was impaired,the glucose tolerance of the body would decrease and then lead to diabetes.Because the clinical manifestation of this disease was not very obvious in the early stage,and the routine method of measuring insulin was cumbersome and expensive,patients rarely got more accurate detection and early warning in the early stage.Therefore,it was of great clinical significance to study a human insulin evaluation algorithm that could be applied to portable non-invasive devices for the early diagnosis and insulin evaluation in the treatment of diabetes patients.Based on the mechanism of noninvasive blood glucose measurement developed by the team in the early stage,this paper made an in-depth study on the evaluation of insulin and noninvasive measurement.Firstly,accorded to the theory of glucose insulin dynamics,a mathematical model fitting the evaluation and measurement indexes of insulin was established,and the noninvasive detection of insulin was realized by using dynamic blood glucose and glycosylated hemoglobin.In this study,support vector machine fitting algorithm model and back propagation neural network algorithm model suitable for small samples were established,and three optimization algorithms were used to optimize the parameters of support vector machine according to the characteristics of support vector machine algorithm.Secondly,through the analysis,the glycosylated hemoglobin with low detection frequency and the oral glucose tolerance test data obtained by noninvasive detection were selected as the input parameters;The output parameter was the evaluation index of insulin.According to the relevant theory and the theory of glucose insulin dynamics,Homeostasis Model Assessment,Insulin Sensitivity index and Modified β-cell Function Index are selected in the steady-state and dynamic relationship of glucose insulin β-cell function index was used as the output of the model.These output parameters could not only better reflect the health level of islet β-cells,and the fasting insulin concentration of the tested object could be obtained through reverse operation.Moreover,the diabetes patients in the Department of endocrinology of our university and the cooperative unit were selected on the research subjects.The blood samples were collected,and the raw data were obtained by enzyme linked immunosorbent assay and glucose oxidase method.After processing,the obtained data were sent to the algorithm model for multiple training and parameter adjustment,and the results obtained from different algorithm models were compared to obtain the algorithm model with the best performance of each evaluation parameter.Finally,through comparison,it could be found that the support vector machine algorithm optimized by the optimization algorithm had better output performance than the back-propagation neural network algorithm,and the correlation of each index was about 80% or more.Among them,the correlation coefficient of Insulin Sensitivity index reached 94.23%,which had more obvious advantages than the back-propagation neural network algorithm model.It could be concluded that it was feasible to use the support vector machine algorithm model optimized by the optimization algorithm to improve the noninvasive detection of insulin evaluation indexes and realize the noninvasive detection of insulin.