Supervised Learning for Non-Invasive Pre-Diabetes, Type 1 and Type 2 Diabetes Screening

In this research we present preliminary work using data extracted from the 2014 Behavioral Risk Factor Surveillance System data published by The Center for Disease Control and Prevention. Through the statistics provided by the Center for Disease Control and Prevention (CDC) on both type 1 and type 2 diabetes, it is noticeable there is a significant rise in diabetic cases. Diabetes is a persistent medical condition that disrupts the body’s insulin processing, resulting in elevated sugar levels that may lead to a range of health complications over the long term. It can affect individuals socially by altering their quality of life, creating an economic impact through the need for insulin and doctor appointments. Hence, the ability to predict the development of diabetes would facilitate early diagnosis and intervention. The main goal is to screen for pre-diabetes, type 1, and type 2 cases and build a high-risk identification system. This work presents a preliminary tackle to screen for type 1, type 2, pre-diabetes, and no diabetes through a comparative analysis of the performance of various trained models using supervised learning in relation to different sampling techniques, including Random Oversampling, Random Undersampling, Synthetic Minority Oversampling (SMOTE), and Adaptive Synthetic Sampling (ADASYN). Our analysis indicates that the K-nearest neighbor algorithm performs best for Type 2 prediction and Naive Bayes for pre-diabetes and Type 1.