Optimizing Feature Selection for Multi-Task Prediction of Non-Communicable Diseases on Imbalanced Healthcare Data

Abstract

This research aims to investigate and compare the performance of feature selection techniques, encompassing Filter and Wrapper methods, for multi-task healthcare data characterized by high class imbalance. The primary objective is to develop a predictive model for non-communicable diseases (NCDs) by integrating the most efficient feature selection strategies with machine learning algorithms, evaluated through statistical metrics appropriate for imbalanced datasets. The methodology consists of six key stages: 1) data collection from Kaggle, comprising 253,680 records and 19 features; 2) data preprocessing and multi-target definition for three diseases; 3) feature selection using Filter methods (Information Gain, Gain Ratio, and Chi-Square) and Wrapper methods (Forward Selection and Backward Elimination); 4) class imbalance mitigation via the SMOTE technique; 5) classification model development using XGBoost, Random Forest, and LightGBM; and 6) model evaluation and hyperparameter optimization to identify the most suitable values. The experimental results reveal that the best-performing model, utilizing the Gain Ratio technique, significantly enhances the performance of the XGBoost model, successfully reducing the feature set from 19 to 7 key features—including BMI, Age, Income, Physical Health, General Health, Education, and Mental Health—while achieving a peak accuracy of 0.8768 and a maximum AUC of 0.9300. These findings indicate that the proposed approach is highly suitable for accurate and efficient multi-task prediction of non-communicable diseases.