Exploratory study on classification of diabetes mellitus through a combined Random Forest Classifier

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Public Health

Background:

• Diabetes Mellitus (DM) is a leading chronic non-communicable disease globally.
• Identifying high-risk individuals is essential for effective DM prevention strategies.
• Medical data often presents challenges like high dimensionality, redundancy, and imbalance.

Purpose of the Study:

• To explore supervised classifiers for DM risk prediction.
• To address challenges of high-dimensional, imbalanced medical data.
• To identify optimal methods for classifying individuals at high risk for DM.

Main Methods:

• Utilized Support Vector Machine-Synthetic Minority Over-sampling Technique (SVM-SMOTE) for data balancing.
• Employed Logistic stepwise regression and Least Absolute Shrinkage and Selection Operator (LASSO) for feature reduction.
• Evaluated four supervised classifiers using Accuracy, Precision, Recall, F1-Score, and Area Under the Curve (AUC).

Main Results:

• The Random Forest Classifier, combined with SVM-SMOTE and LASSO, demonstrated superior performance (Accuracy=0.890, AUC=0.948).
• This combined approach significantly enhanced classification for predicting high-risk DM individuals.
• Key predictive variables identified include age, region, heart rate, hypertension, hyperlipidemia, and BMI.

Conclusions:

• The Random Forest Classifier with SVM-SMOTE and LASSO is highly effective for identifying individuals at high risk of DM.
• This combined methodology serves as a valuable tool for early DM screening.
• The study highlights critical factors influencing diabetes risk, aiding targeted interventions.