Abstract
OBJECTIVE: To predict the two year risk of respiratory admission to hospital among individuals with chronic obstructive pulmonary disease (COPD), with the development and validation (internal and external) of a prognostic score. DESIGN: Model development and validation in population cohorts. SETTING: Birmingham Lung Improvement Studies (BLISS) cohort of new and existing patients with COPD in primary care (model development and internal validation); Evaluation of COPD Longitudinally to Identify Predictive Surrogate Endpoints (ECLIPSE) international cohort and UK primary care Clinical Practice Research Datalink (CPRD) Aurum database linked with Hospital Episode Statistics (external validation). PARTICIPANTS: 1894 patients with new and existing COPD from BLISS cohort; 1749 patients with moderate to very severe COPD from ECLIPSE cohort; 27 340 patients with COPD from CPRD Aurum database linked with Hospital Episode Statistics. MAIN OUTCOME MEASURES: One or more respiratory admissions within two years of cohort entry for development, internal validation, and external validation in CPRD; severe exacerbation within two years for external validation in ECLIPSE cohort. The model was developed from 23 candidate predictors by using multivariable logistic regression with bootstrapping for internal validation and adjustment for overfitting and optimism. Discrimination and calibration were assessed at each stage. Net benefit of the score (clinical utility) was examined across a range clinically relevant risk thresholds compared with use of individual score components. Subgroup and sensitivity analyses were conducted in the CPRD. The BLISS score was directly compared with the Bertens' score in the ECLIPSE cohort. Clinical implementation was explored with relevant stakeholders. RESULTS: Six predictors were retained (age, COPD Assessment Test score, respiratory admissions in the previous 12 months, body mass index, diabetes, forced expiratory volume in 1 second % predicted) to form the BLISS score for estimating an individual's two year risk of respiratory admission. The score had similar discrimination performance on internal validation (optimism adjusted C statistic 0.73 (95% confidence interval 0.70 to 0.77)) and external validation (ECLIPSE: C=0.73 (0.71 to 0.76); CPRD: C=0.71 (0.70 to 0.72)) and good calibration performance in the BLISS (slope=0.87 (95% confidence interval 0.73 to 1.02), CPRD (0.89 (0.85 to 0.93)), and ECLIPSE (0.92 (0.79 to 1.05) cohorts). Stratified analysis in the CPRD cohort showed that it was robust in different population subgroups. Net benefit analyses showed superiority of the BLISS score over individual predictors and the Bertens' score (C=0.68 (0.65 to 0.71); calibration slope 0.68 (0.56 to 0.81)). CONCLUSIONS: The BLISS score showed good performance in estimating individual risk of respiratory admission (within two years) in cohorts containing patients from different settings and geographical locations and with different severities of COPD. Four of the included six variables are readily available in primary care records, and two are partially available but easy to collect. Impact evaluations are now needed to fully study use of the score in clinical care. STUDY REGISTRATION: ECLIPSE ClinicalTrials.gov NCT00292552.