A Machine Learning Approach for Pulse Detection using Cerebral Oximetry Signals during Out-of-Hospital Cardiac Arrest

Out-of-hospital cardiac arrest (OHCA) is a leading cause of death worldwide, and detecting the return of spontaneous circulation (ROSC), which essentially relies on the discrimination between pulseless electrical activity (PEA) and pulse-generating rhythm (PR), is critical for patient survival. ECG-based automatic PEA/PR classifiers proposed to date, often show poor performance due to the significant similarities both rhythms present in the ECG pattern. The high temporal resolution cerebral oximetry signal obtained through near-infrared spectroscopy reflects fluctuations correlated with the QRS complexes when a PR rhythm is present. This study introduces a novel classifier based on a random forest (RF), which combines cerebral oximetry and ECG-derived features to improve PEA/PR discrimination. The study dataset included 165 5-second segments of concurrent ECG and oximetry signals, with 90 PR (8 patients) and 75 PEA segments (12 patients). These segments were randomly and patient-wise stratified into a 5-fold cross-validation (CV) for feature selection and RF training/evaluation. A initial set of 350 PEA/PR discrimination features was evaluated, and using a combination of area under the curve scores and recursive feature elimination, the feature set was reduced. In order to assess the added value of cerebral oximetry in PEA/PR discrimination, the proposed classifier was also modeled using solely ECG-based features. The results show that the combined model achieves similar performance to the ECG-only approach, but using substantially fewer features. The combined classifier obtained a mean (standard desviation) specificity, sensitivity and balanced accuracy of 88.3 (5.1)%, 82.4 (4.8)% and 85.3 (2.8)%, respectively, across 20 replicas of the 5-fold CV when only 11 features were used. These findings suggest that the inclusion of high-resolution oximetry signals significantly enhances the detection of ROSC, offering a more computationally efficient and interpretable approach compared to traditional ECG-based methods during OHCA.