Calibration and performance evaluation of PM2.5 and NO2 air quality sensors for environmental epidemiology

Over the past few decades, the study and the use of air quality sensors have significantly increased, leading to a wealth of experience and a deeper understanding of their strengths and limitations. This study aimed to develop and evaluate a methodology for PM2.5 and NO2 sensors to enhance sensor accuracy to a level suitable for epidemiological studies, where ensuring data quality is paramount. The performance evaluation of indoor and outdoor sensors was carried out during the co-location phase with reference-equivalent instruments (RIs), by calculating the relative expanded uncertainties (REUs) stated in the EU Air Quality Directive 2008/50/EC and the recently published EU Directive 2024/2881, target diagrams and common error metrics, before the deployment of the air quality sensor systems (AQSSs) in the houses of patients suffering from chronic obstructive pulmonary disease (COPD) or asthma, in Stuttgart (Germany). Regression and machine learning (ML) models for sensor calibration were tested during the co-location. Moreover, an original methodology was designed and evaluated to validate the sensor data during the deployment in the houses of the participants. The study found that indoor sensor calibration using artificially generated NO2 and aerosols does not ensure model transferability, emphasizing the need for training data that matches the intended deployment environment in terms of real patterns of concentration, particle composition and environmental conditions. Moreover, the effect of the aggregation time (1, 5, 10, and 15 min) on the performance of the calibration models was evaluated for NO2 sensors. Integrating metadata such as activity logs, window status, and data from official monitoring stations, as well as NO2 measurements with diffusion tubes, proved to be helpful for data validation and interpretation during the sensor deployment in the houses of the participants.