Real-time detection of return of spontaneous circulation during cardiopulmonary resuscitation using AI and carotid Doppler ultrasound

Manual pulse palpation during cardiopulmonary resuscitation (CPR) is often unreliable, potentially delaying the timely detection of return of spontaneous circulation (ROSC). To address this, a novel continuous, hands-free carotid Doppler ultrasound system, RescueDoppler, has been developed to detect spontaneous circulation during CPR. However, interpreting Doppler blood flow patterns in high-stress emergency settings can be challenging for rescuers. To support clinical decision-making, we have integrated state-of-the-art deep learning algorithms capable of automatically identifying the presence of ROSC and spontaneous circulation during chest compressions.

Employing advanced deep learning techniques to automatically detect signs of spontaneous circulation and provide real-time feedback on signs of spontaneous circulation during CPR and ROSC.

We retrospectively analysed carotid blood flow velocity data from five cardiac arrest patients using the RescueDoppler system during CPR. A total of 2608 heart cycles were annotated as either chest compressions with spontaneous circulation or ROSC (Figure 1). One-second pulse segments were used to train a model based on ResNet101 (SOTA convolutional network) for feature extraction, followed by a neural network classifier to distinguish the signals. A leave-one-subject-out approach was used, with the test subject exhibiting both classes, including ROSC. To interpret model’s predictions, we applied Gradient-weighted Class Activation Mapping (Grad-CAM) to visualize the model’s focus areas (Figure 2).

We included 4 men and 1 female with cardiac arrest (67 ± 11 years). The causes of arrest included cardiac events (n = 3), and septic shock (n = 2). Initial rhythms were pulseless electrical activity (PEA, n = 2), ventricular fibrillation (n = 2), and asystole (n = 1). Our model demonstrated strong performance, achieving a mean sensitivity of 95%, specificity of 95%, positive predictive value (PPV) of 95%, and negative predictive value (NPV) of 95%. These high PPV and NPV values reflect the model’s reliability in distinguishing between ROSC and spontaneous circulation during compressions. Explainable AI (XAI) heatmaps revealed that the model consistently focused on physiologically relevant features—such as baseline noise and spontaneous circulation peaks for identifying compressions with spontaneous circulation (Figure. 2a) and the dicrotic notch and diastolic flow for ROSC detection (Figure. 2b).

Deep learning models integrated with real-time Doppler ultrasound can accurately detect spontaneous circulation during chest compressions and ROSC during CPR, even in real time with appropriate Graphics Processing Unit support. This demonstrates the potential of AI to deliver immediate feedback during cardiac arrest, enhancing clinical decision-making and potentially improving outcomes.Figure 1:

Annotations