Proxy-based modeling of indoor air pollution exposure and nonlinear health patterns in biomass-dependent households using machine learning.

This study seeks to address critical gaps in household air pollution (HAP) exposure assessment in low- and middle-income settings by applying a proxy-based exposure reconstruction framework supported by machine learning to develop an exploratory proxy-based exposure index and examine associations with nonlinear health patterns in Kabale, Uganda. A quantitative cross-sectional survey (N = 275) was conducted to capture behavioral, environmental, and socio-demographic proxies of exposure, including fuel type, cooking duration, ventilation, and cooking location. A latent exposure index was developed using Principal Component Analysis and validated with Random Forest and Gradient Boosting models, while nonlinear exposure-health relationships were analyzed using ML classifiers and household typologies identified through K-Means clustering. The findings reveal that exposure is driven primarily by behavioral factors, particularly cooking duration, rather than fuel type alone. Nonlinear modeling suggested possible threshold-like and plateauing patterns in self-reported health outcomes, with acute symptoms such as eye irritation appearing at lower proxy-based exposure levels and chronic respiratory distress showing possible plateauing at higher exposure levels. Clustering analysis identifies a large high-risk group characterized by prolonged indoor biomass use and elevated symptom burden. ML models outperform traditional regression approaches in predicting chronic symptoms but show limited performance for clinical diagnoses. These findings suggest that proxy-based HAP exposure patterns may be multidimensional and potentially nonlinear and demonstrate the value of ML in proxy-based environments, emphasizing that policy interventions must move beyond fuel switching to address behavioral practices and structural constraints shaping exposure.