Robust identification of environmental exposures and community characteristics predictive of rapid lung disease progression.

ABSTRACT

Environmental exposures and community characteristics have been linked to accelerated lung function decline in people with cystic fibrosis (CF), but geomarkers, the measurements of these exposures, have not been comprehensively evaluated in a single study. To determine which geomarkers have the greatest predictive potential for lung function decline and pulmonary exacerbation (PEx), a retrospective longitudinal cohort study was performed using novel Bayesian joint covariate selection methods, which were compared with respect to PEx predictive accuracy. Non-stationary Gaussian linear mixed effects models were fitted to data from 151 CF patients aged 6-20 receiving care at a CF Center in the midwestern US (2007-2017). The outcome was forced expiratory volume in 1 s of percent predicted (FEV1pp). Target functions were used to predict PEx from established criteria. Covariates included 11 routinely collected clinical/demographic characteristics and 45 geomarkers comprising 8 categories. Unique covariate selections via four Bayesian penalized regression models (elastic-net, adaptive lasso, ridge, and lasso) were evaluated at both 95 % and 90 % credible intervals (CIs). Resultant models included one to 6 geomarkers (air temperature, percentage of tertiary roads outside urban areas, percentage of impervious nonroad outside urban areas, fine atmospheric particulate matter, fraction achieving high school graduation, and motor vehicle theft) representing weather, impervious descriptor, air pollution, socioeconomic status, and crime categories. Adaptive lasso had the lowest information criteria. For PEx predictive accuracy, covariate selection from the 95 % CI elastic-net had the highest area under the receiver-operating characteristic curve (mean ± standard deviation; 0.780 ± 0.026) along with the 95 % CI ridge and lasso methods (0.780 ± 0.027). The 95 % CI elastic-net had the highest sensitivity (0.773 ± 0.083) while the 95 % CI adaptive lasso had the highest specificity (0.691 ± 0.087), suggesting the need for different geomarker sets depending on monitoring goals. Surveillance of certain geomarkers embedded in prediction algorithms can be used in real-time warning systems for PEx onset.