Development and Preliminary Validation of an Explainable Machine Learning Model for Early ROSC Detection during Chest Compression in a Swine Model of Cardiac Arrest

• An explainable ML model for ROSC detection during CPR was built with non-invasive ECG, ETCO2 and PPG in CA mini-swine models. • Eight ML algorithms were compared; AdaBoost performed best (AUROC 0.891, sensitivity 0.813, specificity 0.957, 18.2s early detection). • SHAP analysis identified ECG high-frequency energy, mean slope and HR as key predictive features for ROSC. • The model avoids CPR interruption, minimizes false ROSC detection, and provides a proof-of-concept for clinical use. This study aimed to develop a novel multi-parameter-based system for the early and accurate identification of restoration of spontaneous circulation (ROSC) during chest compression using a machine learning (ML) approach in cardiac arrest (CA) animal models. Thirteen CA mini-swine models were established by inducing ventricular fibrillation, followed by cardiopulmonary resuscitation. Electrocardiograph (ECG), end-tidal carbon dioxide (ETCO2), and photoplethysmography (PPG) waves were recorded during and after resuscitation. 75 non-ROSC segments (24276 data points, each second generates a data point) and 52 ROSC segments (1416 data points) were recorded. Nine animal datasets were assigned to the training set, and four to the test set. Eight ML methods, including random forest (RF), support vector machine (SVM), decision trees (DT), neural networks (NNET), Naive Bayes (NB), Nearest neighbor (KNN), logistic regression (LR), and AdaBoost, were used to construct the model. Model performance was evaluated using multiple metrics, including the area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, positive predictive value (PPV), F1-score, and leading time. The SHapley Additive exPlanation (SHAP) method was used to rank the feature importance and explain the final model. Among the eight ML models, the Adaboost model performed the best. The Adaboost model accurately identified ROSC without the interruption of chest compression AUROC 0.891, sensitivity 0.813, specificity 0.957, PPV 0.929, F1-score 0.867 and leading time 18.2s. SHAP analysis identified the top dominant predictors of ROSC: ECG high-frequency energy, ECG mean slope, heart rate, respiration rate, perfusion index, and median ETCO2. A machine learning model integrating non-invasive ECG, ETCO2, and PPG waveform parameters achieved reliable ROSC recognition during uninterrupted chest compression in animal models, with a detection time 18.2 seconds earlier than clinical judgment. The proposed model holds potential for future clinical translation and warrants further external validation in larger clinical studies.