Optimization of Diabetes Mellitus Classification Using the Random Forest and SMOTE-ENN Methods
TL;DR
The Random Forest algorithm, when backed by suitable data preprocessing, is very efficient and could be utilized in clinical decision support systems as an early tool for diabetes screening.
Abstract
Diabetes Mellitus is a non-communicable disease (NCD) that has turned into a worldwide health issue with a steadily rising prevalence. Timely identification is essential for minimizing the risk of complications and the financial strain on the healthcare system. This research focuses on creating a precise and dependable diabetes classification model through the Random Forest algorithm by implementing a series of systematic data preprocessing methods. This methodology utilizes a dataset obtained from Kaggle, consisting of 768 samples. The steps taken include addressing missing values using Multiple Imputation by Chained Equations (MICE) , removing outliers with the Z-Score and Interquartile Range (IQR) techniques , selecting features based on ANOVA F-value to identify the eight most significant features , and balancing classes through the Synthetic Minority Over-sampling Technique-Edited Nearest Neighbours (SMOTE-ENN) to correct dataset imbalance. The assessment of the Random Forest model revealed outstanding performance, attaining an accuracy of 94.3% and an Area Under Curve (AUC) value of 0.98. These findings suggest that the model possesses strong discriminative capability to differentiate between diabetic and non-diabetic individuals. This research concludes that the Random Forest algorithm, when backed by suitable data preprocessing, is very efficient and could be utilized in clinical decision support systems as an early tool for diabetes screening.