Research Of Chronic Disease Prediction Technology Based On Machine Learning

In the field of medical and health,it is very important to diagnose and predict the condition of patients with chronic diseases.After hospital treatment,it is easy for patients with chronic diseases to recur in a short period of time,while the condition can be more serious.This will bring a huge physical and psychological blow to patients as the burden increasing,meanwhile valuable medical resources will be making away.With the development of artificial intelligence technology,it is possible to predict chronic diseases.Based on the physiological information of patients,disease history,doctor diagnosis information and local medical conditions and other factors,the use of machine learning model,can predict the future condition of patients.In this thesis,several diabetes prediction models based on machine learning are constructed based on the inpatient data set of diabetic patients.The main research work and achievement:(1)The Factorization Machines,is studied,which can effectively solve the problem that polynomial model is difficult to train in sparse data.The Factorization Machines is implemented by using the deep learning framework and fully connected layer,while the experiment is added to the Factorization Machines.The improved Factorization Machines can excavate the high-order feature combination information,thus improving the accuracy of the model.(2)A feature embedding algorithm based on Factorization Machines is proposed.The improved Factorization Machines is applied to the unsupervised training between features.The output of the model embedding layer is used as the feature embedding of the original class features.The experimental results show that the features of these embedded features are of high importance and can improve the accuracy of the model.(3)The class feature frequency crossover matrix is studied,and the method of feature crossover using category feature frequency crossover matrix is summarized.This method not only improves the operation speed of feature extraction,but also facilitates the parallelization of feature extraction.Then a new feature embedding algorithm based on class feature frequency crossover matrix is proposed.The experimental results show that the embedded features obtained by this algorithm are of high importance and can improve the accuracy of the model.Finally,a diabetes prediction model based on feature embedding algorithm is constructed by using theabove algorithms.(4)Four diabetes prediction models were constructed by using gradient lifting decision tree,support vector machine,random forest and deep learning technology.After that,the related techniques of model fusion are studied,and the above four diabetes prediction models are integrated by linear weighted fusion and stacking fusion algorithm respectively.The experimental results show that the effect of the model fusion is better than that of the sub-model.Finally,the stacking algorithm is improved,and the weak classifier in the sub-model is directly used in the model fusion.The experimental results show that the improvement can improve the accuracy of the fused model.