Resuscitation: Artificial Intelligence-Assisted, Electrocardiogram-Derived, Sequential RuleBased Decision Tree for Exclusion of Acute Coronary Syndrome as the Etiol‐ ogy of Out-of-Hospital Cardiac Arrest

Background: After out-of-hospital cardiac arrest (OHCA) with return of spontaneous circulation (ROSC), deciding which patients without ST-segment elevation warrant emergency coronary angiography is difficult, as standard electrocardiogram (ECG), troponin, and echocardiography are limited in the early post-arrest window. We aimed to develop a sequential rule-based decision tree to identify post-ROSC patients in whom acute coronary syndrome (ACS) is unlikely to be the arrest etiology.

Methods: In this single-center retrospective study, adult patients with OHCA were divided into development (2020-2024; n=329) and temporal validation (2025; n=71) cohorts. Post-ROSC 12-lead ECGs were analyzed by a deep-learning platform generating ten digital biomarkers. Five candidate biomarkers were combined into a rhythm-stratified sequential decision tree under a zero false-negative constraint; its performance was compared with that of individual biomarkers and validated in the temporal cohort.

Results: The tree classified 171/329 (52.0%) and 32/71 (45.1%) patients in the development and validation cohorts, respectively, as "ACS not likely," with no adjudicated ACS case misclassified at any node. Yield concentrated in the non-shockable subgroup, in which the model excluded 169/261 (64.8%) non-ACS cases (area under the receiver operating characteristic curve [AUROC], 0.824), outperforming each individual biomarker. In the shockable subgroup, which carries a high pretest probability of ACS, 38/40 patients (95.0%) were retained as "ACS not excluded."

Conclusions: A rhythm-stratified, rule-based decision tree combining AI-assisted ECG-derived biomarkers classified approximately half of post-OHCA patients as "ACS not likely" with no adjudicated ACS case misclassified, principally in the non-shockable subgroup. This transparent, bedside-deployable framework warrants prospective multicenter evaluation.