Machine Learning-Based Return of Spontaneous Circulation Prediction During Cardiopulmonary Resuscitation in a Swine Model of Cardiac Arrest: Effect of Data Resolution and Multimodal Physiological Waveforms

Objectives: To determine whether high-resolution (HighRes) and multimodal integration of physiologic signals improve prediction of return of spontaneous circulation (ROSC) during pediatric cardiopulmonary resuscitation (CPR) compared with low-resolution (LowRes) and single-modality approaches. Design: Retrospective analysis of experimental data using machine learning models for outcome prediction. Setting: Laboratory setting with pediatric swine models of cardiac arrest. Subjects: A total of 187 pediatric swine undergoing standardized cardiac arrest and CPR protocols. Interventions: Animals were monitored using multiple physiologic signals during CPR, including aortic blood pressure (ABP), right atrial pressure (RAP), capnography, and electrocardiography. No therapeutic interventions were evaluated. Measurements and Main Results: Four data approaches were evaluated: 1) Waveform-HighRes (100 Hz waveforms); 2) Compression-HighRes (compression-by-compression physiologic series); 3) Waveform-LowRes (15-s averaged waveforms); and 4) Compression-LowRes (15-s averaged compression-by-compression series). Models were developed to predict ROSC using segments 2–4, 2–6, 2–8, and 2–10 minutes of CPR, using both single and combined signal modalities. Area under the receiver operating characteristic curve (AUROC) was used to evaluate models’ performance. In early CPR (2–4 min), Compression-HighRes outperformed both LowRes approaches for ABP (AUROC, 0.74 0.65–0.82 vs. 0.65 0.55–0.74 and 0.54 0.44–0.64) and RAP (0.70 0.62–0.79 vs. 0.61 0.51–0.70 and 0.57 0.48–0.66; p < 0.05). In multimodal models, LowRes data performed comparably to HighRes models (AUROC, 0.76–0.79). Across time points, ABP-based model performance improved, reaching AUROC 0.90 (0.84–0.95) for the full CPR period (2–10 min)—comparable to the multimodal model (0.89 0.83–0.95). Conclusions: HighRes monitoring improved early ROSC prediction for individual signals, especially ABP and RAP. However, combining multiple modalities compensates for lower resolution, enabling comparable predictive performance. These findings support data-driven strategies for selecting physiologic targets and technical requirements in physiology-directed CPR.