Evaluating and Comparing Machine Learning Models for PM2.5 Health Impact Assessment in Thailand

Abstract

This study presents a methodological comparison of four machine learning and statistical models to assess the acute impact of fine particulate matter (PM2.5) on public health in Northern Thailand. The performance of Random Forest (RF) and Long Short-Term Memory (LSTM) was evaluated against two baselines: Linear Regression (LR) and SARIMAX. Air quality data and health surveillance data from 2023 were integrated for four disease groups in Health Region 2. The final merged datasets comprised approximately 5 0 0 - 6 0 0 provinceday records per disease group. The results demonstrate that Random Forest was the most robust model, achieving the highest R -squared scores for Skin (R²=0.21) and Eye diseases (R² = 0. 1 8). The LSTM model showed competitive performance, ranking second, whereas Linear Regression failed to capture the non-linear patterns. Conversely, all models yielded negative R² values for Cardiovascular disease, suggesting that short-term exposure (2-day lag) is insufficient for predicting heart-related conditions.