Heterogeneous effects of alveolar recruitment in acute respiratory distress syndrome: a machine learning reanalysis of the Alveolar Recruitment for Acute Respiratory Distress Syndrome Trial.

BACKGROUND: Despite a robust physiological rationale, recruitment manoeuvres with PEEP titration were associated with harm in the Alveolar Recruitment for Acute Respiratory Distress Syndrome Trial (ART). We sought to investigate the potential heterogeneity in treatment effects in patients enrolled in the ART, using a machine learning approach.
METHODS: The primary outcome was hospital mortality. Patients were clustered using baseline clinical and physiological data using the k-means for mixed large data method. The heterogeneity in treatment effect between clusters was investigated using Bayesian methods. We further investigated whether baseline driving pressure could modulate the association between treatment arm, cluster, and mortality.
RESULTS: Data from all 1010 patients enrolled in the ART were analysed. Partitioning suggested that three clusters were present in the ART population. The largest cluster (Cluster 1) was characterised by patients with pneumonia and requiring vasopressor support. Recruitment manoeuvres with PEEP titration were associated with higher mortality in Cluster 1 (probability of harm of >98%), but this association was absent in Clusters 2 and 3 (probability of harm of 45% and 68%, respectively). Higher baseline driving pressure was associated with a progressive reduction in the association between alveolar recruitment with PEEP titration and mortality.
CONCLUSIONS: Recruitment manoeuvre with PEEP titration may be harmful in acute respiratory distress syndrome (ARDS) patients with pneumonia or requiring vasopressor support. Driving pressure appears to modulate the association between the ART study intervention, aetiology of ARDS, and mortality. This machine learning approach may help tailor future RCTs. CLINICAL TRIAL REGISTRATION: NCT01374022.