Interpretable machine learning prediction of visits for respiratory diseases in Lanzhou, China: impact of air pollution and meteorological variables

Abstract

Meteorological variables and air pollution have raised the risk of associated mortality or morbidity of diseases, particularly respiratory diseases. In this study, prediction models were constructed employing three machine learning algorithms: generalized additive model (GAM), eXtreme Gradient Boosting (XGBoost) and Random Forest (RF) based on daily meteorological and air quality data, as well as visits for respiratory diseases spanning the years from 2013 to 2018 in Lanzhou, Northwest China. By employing cross-validation and hyper-parameter optimization, we found that the three models demonstrated optimal performance when considering factor inputs with a cumulative lag of 3 days. RF emerged as the most effective predictive model, with a coefficient of determination (R²) of 0.558, a root mean squared error (RMSE) of 16.861, and a mean absolute error (MAE) of 12.854. RF improved over GAM by 21.04 % in terms of R², 8.11 % for RMSE and 13.63 % for MAE. The minimum temperature (T_min) exhibited the highest bias explanatory rate and adjusted R² in GAM model, followed by NO₂ and the maximum temperature (T_max). Results analysis on the RF and XGBoost models using the interpretable SHAP method revealed that O₃ was the most important factors influencing respiratory visits. The distributions of factor importance and their interactions suggested that respiratory risk elevated with the increased pollutant concentrations (NO₂, CO, SO₂) when T_max was below 16 °C or T_min fell below 0 °C.