Predicting hourly indoor ozone concentrations with sensor-based measurements and easily accessible predictors.

Few studies have predicted indoor ozone (O) levels using machine learning methods. This study aimed to predict hourly indoor O concentrations using easily accessible predictors and a machine learning algorithm. We took measurements of indoor O concentrations based on low-cost sensors in 18 cities in China, along with ambient O concentration, meteorological factors, and a binary window status indicator as a proxy for ventilation behaviour, to establish random forest models. The results showed that including window status as a predictor improved model performance, with the cross-validation R increasing from 0.80 to 0.83 and the root mean square error (RMSE) decreasing from 7.89 to 7.21 ​ppb, highlighting the importance of considering ventilation behavior in enhancing model accuracy. The model also effectively captured hourly variations in indoor O, revealing that indoor O concentrations were consistently lower and more stable than outdoor levels. These differences suggest that relying solely on ambient data may misrepresent true personal exposure, underscoring the need to incorporate indoor exposure in assessments. This is the first study to apply easily accessible variables and machine learning methods for indoor O prediction at a large geographic spatial scale, showing promising potential for improving the accuracy of exposure assessments in epidemiological studies.