Humans Versus Machine: Machine Learning‐Based Individualized Treatment Effect Modeling for Extracorporeal Cardiopulmonary Resuscitation in Out‐of‐Hospital Cardiac Arrest Authors: Anna Condella, Nicholas J Johnson Published: 2026 Extracorporeal cardiopulmonary resuscitation (ECPR) continues to grow as a rescue modality for refractory cardiac arrest. When deployed in highly selected patients with out‐of‐hospital cardiac arrest, ECPR improves survival and neurological outcomes. 1 , 2 Yet, patient selection and logistics remain challenging. The majority of prognostic scores developed to guide patient selection for extracorporeal membrane oxygenation lack validation or perform no better than general measures of illness severity, such as the Sequential Organ Failure Assessment score. 3 , 4 Current selection strategies, whether based on trial‐derived inclusion criteria or hospital‐specific protocols, vary widely, reflecting uncertainty in optimal candidate identification and producing inconsistent results. 5 , 6 , 7 In this issue of the Journal of the American Heart Association (JAHA), Chen et al. address this gap by moving beyond static eligibility rules toward a dynamic, data‐driven approach, applying machine learning–based individualized treatment effect (ITE) modeling to determine which patients with out‐of‐hospital cardiac arrest are most likely to derive survival benefit from ECPR. 8 In this 9‐year, multicenter retrospective study across 4 Taiwanese tertiary care centers, the investigators analyzed 1953 adult patients with nontraumatic out‐of‐hospital cardiac arrest matched for witnessed arrest and initial shockable rhythm. They developed a causal forest model using gradient‐boosted trees to estimate patient‐level differences in survival to discharge with versus without ECPR. They aimed to identify subgroups with the highest observed incremental benefit from ECPR. In the top 10% of patients with the highest predicted benefit, survival to discharge reached 50% in the ECPR group compared with 20% in the non‐ECPR group. No trial‐based (ARREST [Advanced Reperfusion Strategies for Patients With Out‐of‐Hospital Cardiac Arrest and Refractory Ventricular Fibrillation], Prague OHCA, INCEPTION [Early Initiation of Extracorporeal Life Support in Refractory Out‐of‐Hospital Cardiac Arrest]) or hospital‐specific rule set produced a statistically significant survival benefit in this data set. Patients in this high‐ITE subgroup were younger on average; had lower pH, lower partial pressure of carbon dioxide, and higher lactate levels; and were more likely to have received bystander CPR, features not universally captured by many ECPR selection criteria. 9 This result underscores a fundamental distinction between prognostic modeling and treatment‐effect estimation. Prognostic models identify patients more likely to survive outright, often those with already favorable conditions, even without the intervention. ITE models, by contrast, seek to identify patients whose survival depends on receiving the intervention. This shift from predicting risk to predicting benefit is particularly important for scarce, high‐cost interventions like ECPR, where the ultimate question is not “Will the patient survive?” but “Will the patient survive because of ECPR?” The analysis also provides unexpected clinical insight: laboratory findings that usually portend poor prognosis, such as severe acidemia and high lactate, emerged here as potential markers of favorable treatment effect. This aligns with the physiologic logic of ECPR, which can restore perfusion and perhaps reverse metabolic derangements more efficiently than conventional CPR. Partial dependence plots suggested that the survival benefit peaks at moderate levels of these derangements (the authors suggest pH 6.95–7.0 and lactate 10–12); patients with extreme values may have surpassed thresholds of reversibility. Compared with stringent rule‐based criteria, the ITE approach also identified more candidates while maintaining high survival rates, potentially broadening access without diluting effectiveness. Access to ECPR remains a vexing challenge. A recent geospatial study in the United States found that only 52% of the adult population live within 30 minutes of an extracorporeal membrane oxygenation capable center by ground transport and that patients with factors such as lower income, higher age, and rural residence were especially likely to face challenges in access. 10 Building more robust systems of ECPR delivery, however, requires accurate modeling of which patients or populations are most likely to benefit clinically from additional resources. With limited overlap among patients selected by different criteria, as the authors' chord diagrams demonstrated, each approach captures distinct subpopulations—a reality that may help explain why interstudy results for ECPR remain inconsistent. Importantly, in this real‐world cohort, only about half of those meeting trial‐based criteria actually received ECPR, illustrating how logistical, resource, and consent factors constrain implementation outside of controlled research environments. Embedding ITE frameworks into real‐time decision support, integrated within electronic health records or bedside applications, could enable rapid, individualized benefit estimation during resuscitation. Thresholds could be adapted dynamically to available resources, using stricter cutoffs in scarcity and broader inclusion when capacity allows. Chen et al. even developed a prototype clinician‐facing web interface, showing how such tools might fit into emergency workflows, at the same time emphasizing it would be used in conjunction with clinical judgment. This study is not without limitations. This retrospective matched analysis cannot eliminate residual confounding. Survival to discharge is not the ultimate goal of cardiac arrest resuscitation; long‐term neurological recovery was not addressed. The highest‐benefit subgroup was small, with wide CIs, and selection thresholds were determined post hoc. Immediate laboratory data such as pH, lactate, and partial pressure of carbon dioxide, pivotal for model performance, may not be universally available in real time, potentially limiting the model's portability to resource‐limited settings. Even in this robust data set, 24.6% of patients were missing a value for lactate. Finally, external validation is essential: the model has undergone only internal testing, and its generalizability across different health systems, patient populations, and ECPR protocols remains unknown. Despite these caveats, Chen et al. provide compelling proof‐of‐concept for precision patient selection for ECPR. By shifting focus from eligibility rules to individualized benefit estimation, their approach may help align clinical decision‐making with the realities of resource scarcity and the ethics of high‐intensity intervention. As machine learning methods mature and point‐of‐care data acquisition accelerates, the integration of ITE models into ECPR workflows could represent a genuine advance: identifying the right patient, at the right time, for the right therapy. This exemplifies a necessary evolution in ECPR decision‐making, one that moves beyond static heuristics toward nuanced, patient‐specific predictions of treatment benefit. In the high‐stakes, time‐compressed arena of refractory cardiac arrest, such tools have the potential not only to improve individual outcomes but also to guide the equitable and efficient allocation of scarce, resource‐intensive therapies like ECPR. The path forward will require rigorous external validation, thoughtful integration into clinical workflows, and careful attention to ethical and logistical considerations. Yet, if these challenges can be met, individualized treatment effect modeling could help transform ECPR from a rescue option of uncertain yield into a strategically deployed intervention, maximizing survival gains while upholding the stewardship of critical care resources. Disclosures Nicholas J. Johnson receives research funding from the National Institutes of Health and the American Heart Association for work unrelated, serves as an unpaid volunteer on guideline committees for both the International Liaison Committee on Resuscitation and the American Heart Association, and serves on advisory boards for Neuroptics, Inc., and Aerogen, Inc. Anna Condella has no disclosures to report. The opinions expressed in this article are not necessarily those of the editors or of the American Heart Association. This article was sent to Sula Mazimba, MD, MPH, Associate Editor, for editorial decision and final disposition. See Article by Chen et al. For Disclosures, see page 3. References