Prediction of the respiratory disease incidence based on environmental factors using machine learning techniques in Penang, Malaysia.

Traditional statistical models often struggle to accurately predict global burden of respiratory diseases due to the complex and interdependent nature of environmental variables. To address these challenges, this study aims to develop and evaluate machine learning models for respiratory disease prediction from 1999 to 2015 in Penang, Malaysia. The dataset on respiratory disease cases was obtained from the Health Informatics Centre, Ministry of Health Malaysia, while environmental data were sourced from the Department of Environment Malaysia. We considered ten predictors for analysis, consisting of air pollution factors such as carbon monoxide (CO), sulphur dioxide (SO2), nitrogen dioxide (NO2), ozone (O3) and particulate matter with a diameter of 10 µm or less (PM10), and climatic variables such as maximum daily temperature (MAXT), minimum daily temperature (MINT), rainfall (R), wind speed (WS) and relative humidity (RH) to identify their influence on respiratory disease. For the prediction models, we implement three main machine learning methods: Decision Tree, Random Forest, and XGBoost. To enhance model performance, we incorporated wrapper methods such as forward selection, stepwise selection, and genetic algorithm in selecting the most relevant features, especially the complexity of environmental data. The model performance was evaluated using RMSE, MAE and NAE in a partitioned with 80/20 percent split. The results showed that the RF model with a genetic algorithm performs better in respiratory disease prediction, achieving RMSE, MAE, and NAE values of 8.4281, 6.6154 and 0.2369, respectively with PM10, SO2, CO, O3, MINT, R, and WS as important features. The SHAP interpretability analysis identified that SO2 is the most influential factor affecting respiratory disease cases, followed by CO, MINT, CO, MINT, PM10, WS, O3 and R.