BACKGROUND: Extracorporeal cardiopulmonary resuscitation (ECPR) has demonstrated survival benefit in selected patients with out-of-hospital cardiac arrest, yet optimal selection criteria remain uncertain. Machine-learning-based individualized treatment effect (ITE) modeling may identify patients most likely to benefit by capturing heterogeneity of treatment response, potentially providing a more accurate strategy for ECPR candidate selection than current rule-based criteria. METHODS: We retrospectively analyzed adult, nontraumatic, emergency medical services-attended patients with out-of-hospital cardiac arrest from 4 tertiary centers in Taiwan between 2016 and 2024. After propensity score matching for shockable rhythm and witnessed arrest, a gradient-boosted trees-based causal forest model was developed to estimate ITE and predict the survival benefit of ECPR to hospital discharge. Absolute observed treatment effect across different ITE thresholds was compared with trial-based (ARREST Advanced reperfusion strategies for patients with out-of-hospital cardiac arrest and refractory ventricular fibrillation trial, PRAGUE OHCA Effect of Intra-arrest Transport, Extracorporeal Cardiopulmonary Resuscitation, and Immediate Invasive Assessment and Treatment on Functional Neurologic Outcome in Refractory Out-of-Hospital Cardiac Arrest trial, and INCEPTION Early Extracorporeal CPR for Refractory Out-of-Hospital Cardiac Arrest trial) and hospital-specific criteria. RESULTS: =0.042). The ITE model identified subgroups with higher observed survival benefit compared with those selected by rule-based criteria. Factors associated with higher ECPR treatment effect included lower serum pH and partial pressure of carbon dioxide level, higher lactate level, younger age, and bystander cardiopulmonary resuscitation. CONCLUSIONS: The ITE-based model was associated with greater observed survival benefit compared with current rule-based criteria and may demonstrate potential as a data-driven framework for candidate selection for ECPR. Further validation in prospective settings is warranted.
Chen et al. (Mon,) studied this question.