A machine learning framework for multimodal temporal prediction of neurological outcome after out-of-hospital cardiac arrest.

Neurological prognostication after out-of-hospital cardiac arrest (OHCA) remains challenging. Existing clinical scores rely on static, single-timepoint assessments and fail to capture the dynamic interplay among coagulation derangement, systemic inflammation, brain injury, and evolving neurological status. Whether integrating serial multimodal data through machine learning can meaningfully improve prediction over established approaches has not been systematically evaluated. We conducted a retrospective cohort study of 414 consecutive OHCA patients treated with targeted temperature management (TTM) at a tertiary cardiac arrest center in South Korea (2009-2021), where withdrawal of life-sustaining treatment is not practiced. Ninety-one features spanning five modalities-coagulation, inflammation, brain injury biomarkers, neurological examination, and static clinical variables-were extracted at admission, 24 h, and 48 h. We compared four machine learning algorithms against single-modality models and a clinical score approximation using five-fold stratified cross-validation. Dynamic prediction models evaluated discriminative performance evolution. SHapley Additive exPlanations (SHAP) analysis quantified feature- and modality-level contributions. Robustness was assessed through temporal validation, self-fulfilling prophecy sensitivity analyses, and exclusion of clinician-decision-dependent variables. Of 414 patients (mean age 55.6 years, 71.8% male), 131 (31.6%) achieved favorable neurological outcome (Cerebral Performance Category [CPC] 1-2) at six months. The full multimodal random forest model achieved an area under the receiver operating characteristic curve (AUROC) of 0.983 (95% CI 0.972-0.991), significantly outperforming the clinical score approximation (AUROC 0.847; ΔAUROC + 0.136, p < 0.001) and every single-modality model (all p < 0.001). At 100% specificity, sensitivity was 0.519. Dynamic prediction improved from AUROC 0.950 at admission to 0.977 at 24 h (p < 0.001) and 0.981 at 48 h. SHAP analysis revealed that neurological examination and brain injury biomarkers dominated overall prediction, while coagulation markers-particularly initial international normalized ratio (INR)-provided the strongest early discriminative signal. The model remained robust on temporal validation (AUROC 0.977), after neurological examination exclusion (0.965), and after excluding clinician-decision-dependent variables (0.979). A multimodal machine learning framework integrating serial thromboinflammation, brain injury, and neurological data substantially outperforms conventional approaches for neurological prognostication after OHCA. The dynamic prediction capability and modality-level explainability offer a pathway toward clinically actionable, time-evolving decision support in post-cardiac arrest care.