Machine learning models for different outcomes to better understand individual treatment response on COPD exacerbations in the IMPACT study

<bold>Aim:</bold> We aimed to develop a machine learning model, trained for time to first COPD exacerbation, to estimate individual treatment response to fluticasone furoate (FF) in addition to umeclidinium/vilanterol (U/V) in the IMPACT trial (<ext-link>NCT02164513</ext-link>), and compared this with a previous model for exacerbation rate. <bold>Methods:</bold> Patients in the FF/U/V (n=3989) and U/V (n=2001) arms were split (70:30) in a train (n=4193) and independent test set (n=1797). Using baseline characteristics, a t-learner of two random survival forests (one for each arm) was trained to predict changes in survival probability in response to added FF. We previously used the same train and test set to predict changes in exacerbation rate (Verstraete, K. et al. Thorax 2023; 78:983-989). <bold>Results:</bold> The survival model (C-index FF/U/V 0.59, U/V 0.63) identified a responder subgroup in the test set: hazard ratio (HR) 0.72 (p=0.01) for the 33% predicted best responders (Q1) and non-significant HR for the other 66% of the cohort (Q2 and Q3). Previously, the rate model found rate ratio (RR) 0.60 (p<0.001) for its Q1 and non-significant RR in Q2 and Q3. Patients in Q1, Q2 and Q3 for rate and survival overlapped partially. Patients predicted to be in Q1 by both models (n=271) were older (p=0.006), had milder disease (higher FEV1, lower SGRQ scores (both p<0.001)) but presented with more previous moderate exacerbations and higher blood eosinophil counts (both p<0.001) compared to patients predicted to be Q3 (33% worst responders) by both models (n=259). <bold>Conclusion:</bold> Combining models for different outcomes may better identify absolute treatment response versus nonresponse on individual patient level.