External validation of lung function predictions in real-world cohorts of COPD

An individualized FEV₁ prediction model has recently been developed based on the Lung Health Study (LHS). An ongoing evaluation of the utility of prediction models in representative settings outside the setup of clinical trials is needed.

To evaluate the external validity of the LHS FEV₁ prediction model in a representative sample of Canadian ever-smoker COPD patients.

We used the sample of ever-smokers with spirometrically-confirmed COPD from the Canadian Cohort of Obstructive Lung Disease (CanCOLD), an ongoing prospective, longitudinal, multicenter population-based cohort study. We compared the mean predicted against the mean observed FEV₁ at follow-up visits. We calculated the root mean square error (RMSE) and the actual coverage probability of the 95% individualized prediction intervals.

The final CanCOLD sample consisted of 360 ever-smoker COPD patients, contributing to 970 FEV₁ observations over an average of 3 years of follow-up (standard deviation (SD): 0.6 years). The mean observed vs. predicted FEV₁ for the first follow-up visit (1.67 years from baseline) was 2.28L vs. 2.28L, and for the second visit (3.13 years from baseline) was 2.19L vs. 2.18L. The RMSE of the predictions was 0.205L, and the actual coverage probability of the 95% prediction intervals was 93%. The model had a similar performance across multiple subgroups.

The model showed robust performance in predicting FEV₁ in the entire sample and in several subgroups of CanCOLD. Our results demonstrate that LHS predictions are valid for at least three years in the general COPD population.