Exacerbations in adults with asthma: A systematic review and external validation of prediction models

Abstract

Background: Several prediction models assessing future risk of exacerbations in adult patients with asthma have been published. Applicability of these models is uncertain because their predictive performance has often not been assessed beyond the population in which they were derived. Objective: This study aimed to identify and critically appraise prediction models for asthma exacerbations and validate them in 2 clinically distinct populations. Methods: PubMed and EMBASE were searched to April 2017 for reports describing adult asthma populations in which multivariable models were constructed to predict exacerbations during any time frame. After critical appraisal, the models’ predictive performances were assessed in a primary and a secondary care population for author-defined exacerbations and for American Thoracic Society/European Respiratory Society–defined severe exacerbations. Results: We found 12 reports from which 24 prediction models were evaluated. Three predictors (previous health care utilization, symptoms, and spirometry values) were retained in most models. Assessment was hampered by suboptimal methodology and reporting, and by differences in exacerbation outcomes. Discrimination (area under the receiver-operating characteristic curve [c-statistic]) of models for author-defined exacerbations was better in the primary care population (mean, 0.71) than in the secondary care population (mean, 0.60) and similar (0.65 and 0.62, respectively) for American Thoracic Society/European Respiratory Society–defined severe exacerbations. Model calibration was generally poor, but consistent between the 2 populations. Conclusions: The preservation of 3 predictors in models derived from variable populations and the fairly consistent predictive properties of most models in 2 distinct validation populations suggest the feasibility of a generalizable model predicting severe exacerbations. Nevertheless, improvement of the models is warranted because predictive performances are below the desired level.