Research On Blood Glucose Prediction Model Based On Respiratory Gases

With the increasing risk of diabetes,the prevention and treatment of diabetes have become medical and social problems.Diabetes is a disease of the body's ability to regulate glucose and does not have the ability to kill.However,the high blood glucose duration will seriously damage the heart,brain,blood vessels,nerves,kidneys,eyes and other important tissues and organs of the human body,resulting in more than 100 kinds of complications.A series of complications of diabetes is easy to occur and difficult to treat.The best treatment is early detection,early intervention.A high blood glucose value is the most intuitive feature of diabetes,so the diagnosis methods are based on the detection of blood glucose.However,the invasive process of blood collection causes the physical pain of the subjects and is not conducive to the prevention of detection.Medical research found that the content of acetone in the respiratory gases of diabetics is much higher than healthy people.A non-invasive blood glucose prediction model was constructed,through analyzing the data of respiratory gases by using the electronic nose,signal processing,machine learning and so on.2330 respiratory samples with blood glucose values collected from hospital were used to construct the blood glucose prediction model.A series of data processing and analysis methods are applied in the model,including: preprocessing by using unified baseline algorithm and standard algorithm;feature extraction by geometry,time frequency analysis and dimension reduction algorithm;reconstruction geometric features by using ReliefF algorithm and improved Mitra-Imp algorithm;feature selection using improved forward selection algorithm and exhaustive method;multi-feature fusion based on feature weights and model scores.Researching on the thousands of samples,the blood glucose prediction model has a stronger ability to summarize the common law.While extracting the peak,slope and integral features of the signals,a new feature extraction method of mean response values of signals was added to fully represent the characteristics of respiratory gases.After expanding the feature set,reorganization of the geometric features avoided the effect of invalid and redundant feature points.The most suitable feature set was generated through the selection of multiple features.A blood glucose prediction model based on multi-feature fusion was constructed,and both the historical performance and the model scores were considered.The experiment results show that the model is feasible to predict the blood glucose levels and has a higher accuracy than the original methods.By comparative analysis,this model is especially suitable for the discrimination of respiratory samples with a normal or abnormal blood glucose value.The model also can be used for the prediction of blood glucose level without hypoglycemia samples.The model is not good for hypoglycemia samples and can be improved in the follow-up work.