An Adoption of a stacking ensemble method to predict noncommunicable diseases in Thailand

Abstract

This study delved into predictive modeling for Non-Communicable Diseases (NCD) prevalence in Thailand, focusing on the significance of Social Determinants of Health (SDH) related features. Through an extensive analysis of various datasets and machine learning models, including Support Vector Regression (SVR), Gradient Boosting Decision Tree (GBDT), Linear Regression (LR), Random Forest (RF), Stacking, and XGBoost, the research evaluated predictive capabilities and explanatory power across different scenarios. Findings revealed the importance of socioeconomic and environmental factors such as household income, air pollution levels, education-related variables, household expenses, and healthcare in predicting NCD occurrence or progression. While SVR occasionally exhibited lower Mean Absolute Error (MAE), it struggled with poor explanatory power, as evidenced by negative or low R-squared and Adjusted R-squared values. Other models, particularly GBDT, RF, and XGBoost, consistently demonstrated superior predictive accuracy and moderate to better explanatory capabilities across various scenarios. The study highlighted challenges including dataset discrepancies, lack of data granularity, and the need for more detailed features, urging future research to address these limitations. Further exploration of additional SDH, incorporation of advanced machine learning techniques, longitudinal studies, and expansion of datasets to include larger and more diverse populations were suggested for improving predictive models' accuracy and explanatory power. These insights offered valuable guidance for healthcare practitioners and policymakers in devising evidence-based strategies to mitigate NCD's impact on public health.