Abstract

Sudden cardiac arrest remains a critical global health issue with survival rates stagnating below 10%. Current resuscitation strategies often employ a “one-size-fits-all” approach, neglecting individual patient variations. Coronary perfusion pressure (CPP), the most reliable indicator of cardiopulmonary resuscitation (CPR) effectiveness, requires invasive catheterization for real-time measurement, limiting its utility in emergency and out-of-hospital scenarios. This study introduces a novel framework for real-time, continuous, non-invasive, and calibration-free CPP estimation, leveraging photoplethysmography (PPG) and electrocardiography (ECG) signals integrated with machine learning. Our method employs an animal-based data-split strategy to enhance generalization and eliminate the need for calibration to new animals, making it highly applicable to real-world situations. A dataset of 13 swine models was collected, each subjected to ventricular fibrillation and resuscitated using mechanical chest compressions aligned with American Heart Association guidelines. During these procedures, solid-state pressure catheters in the aortic arch and right atrium recorded CPP, while PPG and ECG signals were gathered simultaneously. The data was transformed into three different input modalities: single-cycle, rolling-window multi-cycle, and stacked multi-cycle. Among these input modalities, the stacked multi-cycle modality, paired with a transformer model trained using scheduled sampling and utilizing both PPG and ECG signals, achieved the best performance. This configuration yielded a mean absolute error of 6.410 mmHg on an animal-based data split, outperforming previous models. This work highlights the transformative potential of PPG and ECG-based models for non-invasive CPP prediction, enabling personalized and data-driven CPR adjustments. By providing subject-based, real-time feedback, this approach promises to optimize myocardial blood flow, improve resuscitation outcomes and advance the standard of care in cardiac arrest management.