Explainable machine learning for neurological outcome prediction in out-of-hospital cardiac arrest survivors undergoing targeted temperature management: a multi-cohort validation study

Abstract Background Early risk assessment in comatose survivors of out-of-hospital cardiac arrest (OHCA) remains clinically challenging, particularly for patients undergoing targeted temperature management (TTM). This study aimed to develop and externally validate an interpretable machine learning model to predict neurological outcomes in TTM-treated comatose OHCA survivors, leveraging multinational registry data to improve generalizability and early-phase characterization. Methods Data were derived from two multi-center registries: the Korean Hypothermia Network prospective registry (KORHN-pro, n = 1,050) and the Taiwan Network of Targeted Temperature Management for Cardiac Arrest (TIMECARD, n = 393). Adult OHCA patients who remained comatose after return of spontaneous circulation (ROSC) were included. The KORHN-pro dataset was used for model development and internal validation via 10-fold cross-validation, while the TIMECARD registry served as an independent external validation cohort. The primary outcome was a favorable neurological status (Cerebral Performance Category score 1–2) at hospital discharge. Eighteen pre-intervention variables were used to train seven machine learning algorithms. The best-performing model was selected based on discrimination metrics. Model interpretability was evaluated using Shapley Additive exPlanations (SHAP) to examine feature importance, interaction effects, and case-level predictions. Results The eXtreme Gradient Boosting algorithm achieved the highest performance, with an area under the receiver operating characteristic curve of 0.925 in internal and 0.852 in external validation. Key predictive determinants included initial shockable rhythm, time to ROSC, adrenaline dose, and Glasgow Coma Scale motor score. SHAP analysis highlighted synergistic effects among features, particularly between cardiac rhythm and early neurological status, which were further illustrated through case-level explanations in the external validation cohort. Conclusion This study presents an interpretable machine learning model for early neurological stratification in comatose OHCA survivors undergoing TTM. Using multinational registry data, the model demonstrated robust performance across both development and external validation cohorts. By integrating clinically relevant predictors, this approach provides individualized estimates to support early assessment and guide future therapeutic considerations. Notably, this study was not designed to compare different TTM temperature strategies, and results should not be interpreted in that context. The explainable framework is intended to complement clinical evaluation without replacing physician judgment or informing treatment withdrawal decisions.