Abstract Body: Introduction: Timely and accurate detection of return of spontaneous circulation (ROSC) during cardiopulmonary resuscitation (CPR) is crucial to avoid unnecessary—and potentially harmful—chest compressions and to facilitate initiation of post-resuscitation care. However, a reliable tool for real time ROSC identification remains unavailable. To the best of our knowledge, this study presents the first-ever ballistocardiography (BCG) signals recorded during out-of-hospital cardiac arrest (OHCA) and introduces a novel artificial intelligence (AI) algorithm for automated ROSC detection.
Materials and methods: Data were collected from OHCA patients treated by the anesthesiologist-staffed rapid response vehicle in Oslo. Recordings were obtained using the LIFEPAK 15 monitor/defibrillator (Stryker, Kalamazoo, MI, USA) alongside two BCG piezoelectric biosensors (Kopera Norway AS, Stavanger, Norway), which were placed on the skin over the carotid artery and the abdominal aorta. The study dataset contained 60 segments —30 pulseless electrical activity (PEA) and 30 pulse-generating rhythms (PRs) —extracted from the very first six OHCA patients monitored with BCG. Each segment contained ECG, abdominal and carotid BCG signals. The signals were first preprocessed and then adaptively filtered to extract the circulatory-related component (CC) from both abdominal and carotid BCG signals. From each CC, nine waveform features were computed and then averaged. A nested cross-validation architecture was employed to 1) select the optimal feature subset for training a linear discriminant analysis model, and 2) evaluate the model’s performance in terms of sensitivity (SE), specificity (SP), balanced accuracy (BAC, defined as the mean of SE and SP), and accuracy (ACC). This procedure was repeated 25 times to estimate the statistical distributions of the performance metrics.
Results: The mean (standard deviation) duration of PR and PEA segments were 4.9 (0.4) s and 3.7 (0.9) s, respectively. The best classification performance was achieved using three features, yielding a median (interdecile range) SE/SP/BAC/ACC of 100.0 (96.7–100.0)/93.3 (90.0–93.3)/96.7 (93.3–96.7)/96.7 (93.3–96.7)%, respectively.
Conclusions: An accurate and reliable AI-driven ROSC detector was developed using waveform features extracted from CCs of BCG signals. Further validation with larger datasets is necessary to confirm these findings and support future clinical implementation.