Explainable Artificial Intelligence for Prognostic Stratification in Out-of-Hospital Cardiac Arrest Patients Undergoing Extracorporeal Cardiopulmonary Resuscitation

Prognostication in patient with out-of-hospital cardiac arrest (OHCA) underwent extracorporeal cardiopulmonary resuscitation (ECPR) remains challenging due to the complexity of clinical variables. We aimed to develop and interpret artificial intelligence (AI) models for early outcome prediction in OHCA patients treated with ECPR, and to identify clinically meaningful patient subgroups through supervised clustering based on model explanations. We retrospectively analyzed data from the SAVE-J II registry, a multicenter registry of adult OHCA patients treated with ECPR in Japan. We defined and developed prediction models for all-cause death: Cohort 1 included all patients for predicting day 1 outcomes using binary classification models, and Cohort 2 excluded patients who died on day 1 deaths and developed survival models for events from day 2 onward. Models were interpreted using Shapley Additive exPlanations (SHAP), and hierarchical clustering based on SHAP values was performed to stratify patients into prognostic subgroups. In cohort 1 (n=1,624, age 60 IQR [49-68]), 433 (26.7%) all-cause death occurred on day 1, and AI models achieved 0.85 of AUC. In cohort 2 (n=1,191, age 59 IQR [48-67]), 752 (63.1%) all-cause deaths occurred from day 2. AI models achieved a mean of time-dependent AUCs of 0.77. SHAP analysis identified different predictive variables between cohorts. SHAP-based hierarchical clustering revealed patient groups with markedly different prognoses. AI models accurately predicted short-term outcomes in ECPR-treated OHCA patients and revealed temporal shifts in key prognostic factors. SHAP-based clustering enabled meaningful stratification and may support personalized treatment strategies. Can AI models accurately predict all-cause death in patients who underwent ECPR (Extracorporeal cardiopulmonary resuscitation) for OHCA (out-of-hospital cardiac arrest) and can SHAP (Shapley Additive Explanations) values reveal clinically meaningful patient subgroups? AI models accurately predicted all-cause mortality, though less so for bleeding. Landmarking patients at day 1 and interpreting the models with SHAP values revealed differing early and later event characteristics. SHAP-based supervised clustering stratified patients into prognostically distinct groups. By employing interpretable AI models, patient prognoses can be estimated while elucidating the underlying factors. AI models will help clinicians make treatment decisions for patients who underwent ECPR.