A STUDY FOR DIABETES PREDICTION USING HYPERPARAMETER TUNING AND MACHINE LEARNING TECHNIQUES

Abstract

Diabetes is the most frequent metabolic condition that causes excess sugar levels in our blood. Patients with diabetes have a body that can't metabolise insulin adequately or can't make enough insulin. Diabetes is a disease which occurs when the glucose level increases in blood. It is a persistent disease that occurs mainly in two ways: First one is, if the adequate insulin is not produced by the pancreas and the second one, if insulin is not used by the body effectively. Insulin Hormone is responsible for regulating Blood sugar. Diabetes can harm our body parts too like eyes, kidneys, nerve, heart and blood vessels. Therefore Predicting diabetes in the earlier phase is very essential to control the diabetes and to save lives. In this study first we did a survey on the previous studies on this topic and after that we implemented our model on the basis of the survey. Presenting a method of detection by symptoms that the person might observe may motivate the person to seek medical treatment more immediately, concluding in a more precise diagnosis and treatment. In this study, We have taken 35 papers (studies in the time span of 2014-2021) out of which we chose 23 papers for further study based upon our requirements. And then we analysed current research in order to conclude the risk factors for diabetes. This research investigates the accuracy of diabetes prediction. It is concentrated on current advancements that have a significant influence on diabetes diagnosis and detection. and we also see by using which medical information and Machine Learning techniques we can predict better. On the basis of our small survey we concluded that RF is the popular technique used among all techniques and PIMA Indians Dataset is used more frequently. And among all used techniques we found out that boosting and SVM has the highest accuracy. Furthermore, We have used in this research logistic regression, KNN, DT, naive bayes, RF and SVM classifiers. After these techniques we tried to optimise our model. We found that all techniques with optimised models after smote performed very well. and for KNN and RF we achieved the highest accuracy. Then we used hyper parameter tuning to optimise our KNN and RF classifier, and we also tuned XGB. Firstly For RF model using GridsearchCV optimization we achieved accuracy nearly 89.21%. And for XGB after hyper parameter tuning we got an v accuracy of 90.13 percent. Which is the best accuracy among all the ML techniques that we have used. As a result we can say that among all the ML models XGB boost after hyper parameter tuning is performed very well.