DIABETES PREDICTION USING MULTILAYER PERCEPTRON

Abstract

Diabetes mellitus is one of the most popular diseases that causes 1,5 million people to die each year. It is the major cause of blindness, kidney failure, heart attacks, stroke, and lower limb amputation. Being the world’s top 9 severe diseases, diabetes puts a burden on the world’s economy and the healthcare system. There are two types of diabetes. Type 1 is generic and in the vast majority of cases shows up early in life. This study focuses on type 2 diabetes which is around 90% of all diabetes cases, and that can be diagnosed. Early diagnosis of diabetes can prevent serious medical complications. In this literature, we are introducing a framework of diabetes prediction based on a Multilayer Perceptron of only 1 hidden layer and 8 neurons, which is lighter than the state-of-art framework which consists of 3 hidden layers and 144 neurons in total. The grid-search method was used for hyperparameter tuning to

maximize Area Under the ROC Curve (AUC) that was chosen as a performance met- ric. Pima Indian Diabetes Dataset was used to conduct the experiments. The dataset

was preprocessed with outlier rejection, missing values imputation, standardization, data scaling, and feature selection algorithms. K-fold cross-validation technique was used to train/test the classification model. The Multilayer Perceptron model was tuned with various hyperparameters as well as the dynamic learning rate. Finally, the best lightweight MLP model consisting of 1 hidden layer that reaches the AUC of 0.90 was obtained. The model performs as well as the state-of-art but is at least 5 times faster in training and more than 80 times more efficient in terms of memory usage.