Understanding how ozone behaves indoors is vital for assessing human health risks, as people spend most of their time inside. This study developed the first large-scale machine learning model capable of predicting hourly indoor ozone concentrations using easily accessible predictors, including outdoor ozone, meteorological conditions, and window-opening behavior. Ozone is a key air pollutant formed by chemical reactions between nitrogen oxides and volatile organic compounds under sunlight. In 2021, long-term ozone exposure contributed to nearly 490,000 deaths worldwide. Although most exposure assessments rely on outdoor data, people typically spend 70%–90% of their time indoors, where ventilation, indoor sources, and building materials all affect actual ozone levels.
Traditional mechanistic models require detailed indoor parameters that are hard to obtain in large-scale studies, while linear regression models struggle with nonlinear environmental relationships. Due to these limitations, there is an urgent need to develop accurate, scalable models that can predict indoor ozone exposure based on accessible environmental and behavioral data. Researchers from Fudan University and the Chinese Academy of Sciences have built a machine-learning model to predict hourly indoor ozone levels across 18 Chinese cities. The study, published in Eco-Environment & Health on July 9, 2025, used random forest algorithms trained on low-cost sensor measurements combined with meteorological and ventilation data. By comparing two models—with and without window-status information—the researchers demonstrated that including ventilation behavior significantly improved prediction accuracy, marking a major step toward more realistic ozone exposure assessment.
The team collected over 8,200 hours of indoor ozone data using portable electrochemical sensors in 23 households. Predictor variables included outdoor ozone levels from high-resolution random-forest and MERRA-2 datasets, meteorological parameters, and window-opening status recorded manually by volunteers. Two random forest models were compared: one excluding and one including window status. Incorporating window behavior raised cross-validation R² from 0.80 to 0.83 and lowered RMSE from 7.89 to 7.21 ppb. The model accurately captured hourly ozone fluctuations and regional differences, performing better in southern than northern China and in the cold than warm season. Predictor-importance analysis showed surface pressure, temperature, and ambient ozone as dominant factors, with ventilation emerging as a crucial behavioral determinant. Diurnal comparisons revealed that indoor ozone concentrations were 40% lower than outdoor levels during the day, underscoring the buffering effect of indoor environments.
Most exposure studies still rely on outdoor ozone data, but that's only half the story, said Prof. Xia Meng, senior author of the study. Our findings show that ventilation behavior—something as simple as whether a window is open or closed—can change exposure dramatically. By integrating such behavioral data with meteorological information through machine learning, we can finally estimate indoor ozone more precisely at large scales. This will strengthen epidemiological studies and help guide public-health interventions in urban and residential settings. This research introduces a practical, low-cost strategy for predicting indoor ozone exposure in real time across large geographic areas. The model can be integrated into health-risk assessments, smart-home monitoring systems, and public-health surveillance platforms, enabling policymakers and scientists to better understand indoor–outdoor exposure differences. Future work could extend the framework to other pollutants such as fine particulate matter or nitrogen dioxide, incorporate smart sensors for automated window tracking, and expand monitoring to diverse climatic zones. Ultimately, this machine-learning approach bridges environmental modeling with daily life, promoting healthier indoor environments in rapidly urbanizing regions. The study is available at https://doi.org/10.1016/j.eehl.2025.100170.
