Machine learning to identify treatment effect heterogeneity related to ICS and LAMA in the IMPACT trial

A machine learning (ML) model developed on the data of the IMPACT trial (<ext-link>NCT02164513</ext-link>) to estimate individual treatment effects (ITE) of LABA/LAMA/ICS compared to LABA/LAMA identified heterogeneity in the treatment effects due to ICS. Blood eosinophil level was the main predictor, next to years smoked and lung function (Verstraete et al., Thorax 2023). We aimed to validate this model on the unexplored LABA/ICS arm compared to LABA/LAMA arm and hypothesized that the responders from the LABA/ICS arm are characteristically similar to the responders from the LABA/LAMA/ICS arm. Additionally, we performed a heterogeneity analysis on the LABA/ICS vs LABA/LAMA/ICS arms. We used data from subjects in the fluticasone furoate/vilanterol (FF/VI, n=3956) and umeclidinium/vilanterol (UMEC/VI, n=600) arms. ITEs were estimated for all subjects, who were then stratified into quintiles based on estimated ITE. In Q1, a significant rate ratio (RR) of 0.64 (p<0.001) was observed. No significant effects were observed in the other quintiles (Fig A). The characteristics of the subjects in Q1 corresponded to the responders in the FF/UMEC/VI arm: elevated blood eosinophil levels and a lower systolic blood pressure. Additionally, we found no heterogeneity in the added effect of UMEC on top of FF/VI as the RRs were similar for each quintile (Fig B). <bold>Conclusion:</bold> Heterogeneity in treatment effects in COPD is ICS but not LAMA driven. <fig><object-id>erj;64/suppl_68/PA3904/F1</object-id><object-id>F1</object-id><object-id>F1</object-id><graphic></graphic></fig>