Research On The Detection Algorithm Of Expiratory Markers In Diabetic Patients Based On Machine Learning

Diabetes is the world’s third most serious threat to human health.The current diagnosis method of diabetes is venous blood collection,which is an invasive method,which is easy to cause resistance of patients and affect treatment.Expiratory acetone can be used as a marker of breath in patients with diabetes.Compared with large breath detection equipment such as chromatography and mass spectrometer,gas sensor has the characteristics of small size,easy operation and low cost,so the use of gas sensor for breath detection has become a research hotspot.In this thesis,a combination of metal oxide gas sensor array and machine learning algorithm is used to identify the target gas acetone and the interfering gas ethanol qualitatively and quantitatively.At the same time,in order to solve the problem of decreasing recognition ability of models or algorithms based on old sensor data due to sensor replacement,this paper proposes to use transfer learning method to solve the problem of different data distribution,improve the identification accuracy of qualitative model and reduce the error of quantitative model.The main research contents are as follows:(1)According to the expiratory markers and concentration of diabetic patients,appropriate sensors were selected to form a gas sensor array.The CGS-8 intelligent gas-sensitive analysis system was used to collect data of single acetone,ethanol and their mixed gases with concentrations ranging from 1 to 13ppm.The polynomial fitting method was used for baseline reduction of data,and Kernel Principal Component Analysis(KPCA)was used for feature extraction.(2)In terms of qualitative identification of gas,this paper uses Adaptive Boosting(AdaBoost)、eXtreme Gradient Boosting(XGBoost)and Support Vector Machine(SVM)to establish a qualitative identification model of gas,and uses grid search to optimize the parameters.The number of parameters,accuracy,training time and time complexity were compared,and AdaBoost was selected as the qualitative recognition algorithm.Experimental results show that the AdaBoost algorithm has a qualitative identification accuracy of 99.722%.The experimental results show that the accuracy of qualitative identification is 94.55%.(3)In terms of quantitative identification of gas,this paper uses Multivariate Relevance Vector Machine(MVRVM)and SVM regression to establish the quantitative identification model of gases.The Root Mean Square Error(RMSE)and coefficient of determination R~2(R-Square)were compared.MVRVM was selected as the quantitative recognition algorithm.The experimental results show that the RMSE of MVRVM algorithm are 0.027 and 0.030,respectively.The results show that the RMSE of quantitative identification of acetone and ethanol gas are 11.59 and 8.72,respectively.(4)Select a new sensor of the same model and series to replace the old sensor for the experiment.Transfer Component Analysis(TCA)is used to process the data and minimize the difference between the source domain data and the target domain data.Compared with the qualitative and quantitative model without transfer learning.The experimental results showed that the accuracy of qualitative identification of gas increased by at least 3%compared with that without TCA,and the RMSE of quantitative identification decreased by up to 12 times compared with that without TCA.