Application Of Machine Learning Methods To Breath Analysis Of Diabetes

Nowadays,diabetes has become a serious disease threatening the health of all human beings.In recent years,the morbidity and mortality of the patients with diabetes are increasing due to the rapid development of Chinese economy,the phenomenon of aging population and the lifestyle is becoming more westernized.As a chronic disease,the early symptoms of diabetes are not very obvious,and large number of patients realize the diabetes until the serious complications appear.The clinical diagnosis of diabetes is mainly based on blood glucose at present.The conventional method is to conduct blood tests directly.This way is invasive that brings about physical trauma.And it will become a heavy psychological burden to the subject in the long run.Breath analysis are becoming a popular research orientation recently.Clinical study indicated that there are endogenous molecules in human breath that can reflect the status of the relevant metabolism in the body.These molecules can be used as biomarkers of certain diseases and have the potential for disease diagnosis.Among all the current equipments of the breath analysis,the electronic noses are extremely beneficial for the popularization of breath analysis with advantages of high efficiency,low cost,easy operation compared with the large-scale equipments such as chromatograph.Therefore,this thesis improves the traditional electronic noses.Firstly,the thesis adopts many kinds of gas sensors with high sensitivity and high measure precision for the potential markers of diabetes in human breath,then designs a miniaturized gas detection system based on the differentiation of sensor arrays.Next,the thesis conducts a large number of experiments in the patient group and healthy group.Finally,the result of the breath analysis using this system will be validated through the data processing.Because of the current status of electronic noses usually adopt the statistical pattern recognition methods,which need to rely on a large number of sample data as the basis of the algorithm research,but the particularity of the human breath brings about the great difficulty to collect them with a large scale.Therefore,this thesis summarizes a set of data processing methods that suitable for gas signals according to the machine learning theories.The first step is the data preprocessing that can eliminate the useless information to enhance the quality of the data.The next step is the feature extraction because the original data contains large amounts of redundant information.There are three methods of dimensional reduction in this thesis,including the Principle Components Analysis,the Local Linear Embedding and the t-distributed Stochastic Neighbor Embedding.In the end,in addition to the shallow machine learning model based on the statistic theory such as Support Vector Machine and k-Nearest Neighbor,this thesis attempts to adopt the convolutional neural network algorithm to the data processing,which belongs to the deep learning approaches.The deep learning method clearly highlights the importance of feature learning.It can automatically learn the feature values that better characterize the essence of the sample data.The final classification result of this article shows that: the convolutional neural network algorithm is used to classify the sample data,and its accuracy rate is 86.36.%,sensitivity 92.31%,specificity 81.63%,compared with KNN accuracy rate 79.20%,sensitivity 77.44%,specificity 80.63% and SVM accuracy rate 78.28%,sensitivity75.38%,specificity 80.63%,all significantly improved.Therefore,it can be considered that deep learning methods can achieve better classification results.