Uncovering Real-World Implementation of HEPA Portable Air Cleaners in Schools: A Bayesian Latent Class Approach to Improve Exposure Assessment
NASIM ILDIRI, Kevin Josey, Nathaniel Ramirez, Anna Segur, Julia Poje, Bowen Du, Jeffrey Siegel, Thomas Jaenisch, Mark Hernandez, University of Colorado at Boulder
Abstract Number: 556
Working Group: Reducing Aerosol Exposure with Control Technologies and Interventions
Abstract
In real-world settings, air quality interventions are not always implemented or maintained as intended, especially when scaled across many sites. These deviations can compromise exposure classification and bias health outcome evaluations. Portable air cleaners (PACs) are increasingly deployed in schools to reduce airborne exposure risks, but their operation is often inconsistent. To address this, we aim to assess the fidelity of PAC implementation and develop a method to infer usage status using sensor-based indoor air quality (IAQ) data. As part of a state-wide trial, over 10,000 PACs were deployed to K–12 schools in 2023, with each intervention classroom assigned two units. Structured observations were conducted in 512 classrooms across 27 Denver Public Schools to assess real-world PAC usage and adherence to study protocols. We are developing a Bayesian Latent Class Analysis (LCA) model to classify classrooms based on CO2-derived air exchange rates, PM2.5 concentrations, and other relevant IAQ indicators. This unsupervised model aims to infer likely “PAC active” versus “PAC inactive” status without relying on deployment records. Site visits revealed a critical implementation gap: more than a year after large-scale PAC deployment, only 29% of classrooms issued PACs consistently operated them; approximately one-third of teachers left them unplugged or inactive, despite being encouraged and educated to operate them. This misalignment between protocol and practice challenges the validity of any subsequent IAQ–health analyses. Preliminary LCA modeling is underway and shows potential to distinguish clusters of classrooms with IAQ patterns indicative of active filtration versus limited or absent intervention effects. Our approach offers a scalable way to identify actual exposure when protocol adherence is uncertain. This methodology is applicable to studies where sensors or behavioral data supplements incomplete implementation records. Improving exposure classification enhances the credibility of health effect estimates and the effectiveness of policy decisions based on them.