High blood eosinophil counts can predict an enhanced response to inhaled corticosteroids (ICS) in patients with chronic obstructive pulmonary disease (COPD), but additional insight in treatment effect heterogeneity is needed to optimize clinical disease management. We investigated if causal machine learning models can detect heterogeneity in the effect of two ICS-containing therapies on both time to first exacerbation and exacerbation rate, and if this approach can identify additional predictors of ICS response. Baseline characteristics from patients in the fluticasone furoate/umeclidinium/vilanterol (FF/U/V) and U/V arms of the InforMing the PAthway of COPD Treatment (IMPACT) trial (ClinicalTrials.gov identifier NCT02164513) were used to train a causal survival forest to estimate the effect of fluticasone furoate (an ICS) on time to first exacerbation for each patient. Similarly, a causal forest was trained for exacerbation rate. Results were averaged over 100 train and validation sets in a Monte-Carlo cross-validation approach. The analysis was repeated for a comparison of the FF/V and U/V arms. This analysis included 4048 FF/U/V, 4034 FF/V and 2025 U/V patients. Significant p -values for AUTOC (area under the targeting operator characteristic) curve indicated that all models ranked patients well according to individualized treatment and that significant heterogeneity was detected in the effect of ICS on both exacerbation outcomes. Respiratory symptoms, reversibility of airflow limitation and lung function, in addition to eosinophils, were identified as important predictors by all models. The contribution of additional predictors in determining the effect of ICS was most apparent for patients at the lower end of the eosinophil spectrum. This hypothesis-generating secondary analysis of IMPACT detected significant heterogeneity in the effect of ICS and indicated that, in addition to eosinophils, other clinical patient characteristics also contribute to the prediction of ICS response on both time to first exacerbation and exacerbation rate. The causal machine learning approach in this analysis can identify treatment effect predictors and patterns of response, to enhance insight in COPD disease phenotypes and heterogeneous treatment effects.
Huts et al. (Wed,) studied this question.