Closed-loop and Artificial intelligence guided cardiopulmonary resuscitation using real-time physiological signals: a systematic review and meta-analysis

Current cardiopulmonary resuscitation (CPR) includes breaks to check rhythm and pulse, while high-quality CPR requires continual chest compressions. Artificial intelligence (AI)-driven “closed-loop” or “near-closed-loop” systems may evaluate CPR-contaminated signals, guide shock and pulse decisions. According to PRISMA statement 2020, We searched databases from 2010 to 2025 for studies on animal or human cardiac arrest where AI-based or algorithmic tools gave real-time intended guidance during CPR or CPR-dominant analysis windows. Evidence was divided into animal perfusion-targeted closed-loop CPR, human shock-advisory algorithms, and human pulse/perfusing-rhythm detection. We classified shockable and non-shockable instances using a bivariate random-effects model and meta-analyses. Many studies reported performance on many electrocardiogram segments per episode; therefore, aggregated findings should be taken as segment-level diagnostic performance, not episode-level accuracy. Fourteen studies met the inclusion criteria: three porcine closed-loop experiments, eight human shock-advisory studies, two pulse-detection studies, and one simulated individualized strategy. Eight shock-advisory studies demonstrated the pooled sensitivity = 0.93 and the specificity = 0.97. There were substantial variation and a high risk of bias in the patient selection and index test domains. Pulse-detection models worked well during clean pauses but not during compressions. Animal controllers enhanced coronary perfusion pressure, end-tidal carbon dioxide (EtCO₂), or ventricular fibrillation surrogates relative to fixed-depth CPR; however, these effects were transient. AI systems can execute guideline CPR in some datasets, and animal studies confirm perfusion-targeted control’s physiological viability. Closed-loop AI-guided CPR is intriguing yet unvalidated, requiring episode-level clinical studies due to fragmented, varied, and methodologically problematic evidence. PROSPERO CRD420251239485.