Study On Algorithms Of Upgrading Dimension And Ensemble Learning For Diabetes Prediction

With the rapid development of the national economy and the rapid improvement of the national economic level,the people’s living standard and material level have improved by leaps and bounds.The convenience brought by science and technology makes us enjoy all kinds of delicious food at home,and a large number of lightweight labor is replaced by machines.Most people enjoy a high quality of life but neglect to live a healthy life.Unhealthy lifestyle habits such as unhealthy diet,lack of exercise and irregular work habits have led to the development of many diseases.According to the 10 th edition of the International Diabetes Federation（IDF）public report,the number of people with diabetes is rising,with 537 million people living with the disease worldwide in 2021 and 6.7 million adults already dying from diabetes and its complications.The total number of people with diabetes in our country is 149 million,which is a very alarming figure.However the diagnosis of diabetes is a complex process.It usually requires multiple indicators combined with the doctor’s experience and knowledge to judge comprehensively.The body is in a state of high blood sugar for a long time before diabetes is discovered,which damages the patient’s organs and blood vessels,causing various complications.If diabetes is detected early and treated and controlled,the complications of diabetes can be well managed and prevented.This will greatly reduce the risk of diabetes to the patient and family.In recent years,with the rapid development of computer hardware technology,digital medicine and intelligent medicine have been widely used and developed.Many researchers have applied machine learning and deep learning algorithms in the biomedical field,and the predictive diagnosis of diabetes has also received a lot of attention.The main methods for predicting diabetes can be divided into three main categories,（1）optimization algorithm combined with classification algorithm,（2）using Ensemble algorithms,and（3）uses an artificial neural network.These methods improve model performance mainly by trying new learners,or selecting more appropriate algorithms,and combining to form better ensemble models.These methods are feasible to some extent,but they ignore that the data is the key to the final performance of the model.The algorithm can only fit the relationship between data and labels as much as possible by constantly optimizing parameters.Better data will make the algorithm have better performance.Therefore,the focus of our research in this paper is to find the shortcomings of the data itself from the perspective of data by analyzing data and algorithms: The chaotic distribution in the sample space is difficult to divide the decision boundary and divide different samples.In order to solve this problem,Cover theorem is introduced: nonlinear projection of complex pattern classification problem to high-dimensional space will be easier to divide than low-dimensional space.This paper use artificial neural networks to reproject data into higher-dimensional space.In the process of dimension upgrading,the loss function is used to control the projection position of sample points,which makes it easier for the data after dimension upgrading to find the decision boundary.This paper used the enhanced dimension data to train the algorithm models and assembled them into a high-precision stacking integrated model,which reduces the generalization error and increases the stability of the model.Finally,We use several indicators to measure and verify our model,such as accuracy,precision,recall,ROC curve and AUC value.The accuracy of our model is obviously better than that of the algorithm proposed on the same data set in recent years.In terms of precision rate,recall rate and AUC value,the data performance after dimension upgrading is significantly better than that before dimension upgrading.