Environment International (Oct 2024)
Modelling personal temperature exposure using household and outdoor temperature and questionnaire data: Implications for epidemiological studies
Abstract
Non-optimal temperature is a leading risk factor for global disease burden. Most epidemiological studies assessed only outdoor temperature, with important uncertainties on personal exposure misclassification. The CKB-Air study measured personal, household (kitchen and living room), and outdoor temperatures in the summer (MAY-SEP 2017) and winter (NOV 2017-JAN 2018) in 477 participants in China. After data cleaning, ∼88,000 person-hours of data were recorded across each microenvironment. Using multivariable linear regression (MLR) and random forest (RF) models, we identified key predictors and constructed personal temperature exposure prediction models. We used generalised additive mixed effect models to examine the relationships of personal and outdoor temperatures with heart rate. The 24-hour mean (SD) personal and outdoor temperatures were 29.2 (3.8) °C and 27.6 (6.4) °C in summer, and 12.0 (4.0) °C and 7.5 (4.2) °C in winter, respectively. The temperatures across microenvironments were strongly correlated (Spearman’s ρ: 0.86–0.92) in summer. In winter, personal temperature was strongly related to household temperatures (ρ: 0.74–0.79) but poorly related to outdoor temperature (ρ: 0.30). RF algorithm identified household and outdoor temperatures and study date as top predictors of personal temperature exposure for both seasons, and heating-related factors were important in winter. The final MLR and RF models incorporating questionnaire and device data performed satisfactorily in predicting personal exposure in both seasons (R2summer: 0.92; R2winter: 0.68–0.70). We found consistent U-shaped associations between measured and predicted personal temperature exposures and heart rate (lowest at ∼ 14.5 °C), but a weak positive linear association with outdoor temperature. Personal and outdoor temperatures differ substantially winter, but prediction models incorporating household and outdoor temperatures and questionnaire data performed satisfactorily. Exposure misclassification from using outdoor temperature may produce inappropriate epidemiological findings.