# Detection of return of spontaneous circulation during cardiopulmonary resuscitation using continuous carotid artery Doppler blood flow monitored by AI in an animal model

## TL;DR

A new AI-powered device called RescueDoppler can detect when blood flow returns during CPR in an animal model, improving the accuracy of identifying successful resuscitation.

## Contribution

The novel RescueDoppler device with AI enables real-time detection of return of spontaneous circulation during CPR using carotid artery blood flow.

## Key findings

- The model achieved 98% sensitivity and 97% specificity in detecting ROSC during CPR.
- XAI heatmaps identified key features contributing to accurate AI predictions.
- The two-stage classifier successfully distinguished between compression-only and compression with intrinsic cardiac activity.

## Abstract

Manual pulse checks during cardiopulmonary resuscitation (CPR) to confirm return of spontaneous circulation (ROSC) are often unreliable and time- consuming. To address this, a novel RescueDoppler device has been developed, consisting of a small ultrasound probe that attaches to the neck and continuously monitors potential blood flow in the carotid artery.

To provide automatic real-time feedback on ROSC using RescueDoppler carotid blood flow during cardiac arrest by employing advanced deep-learning techniques.

We conducted two experiments using carotid blood flow velocity recordings from 9 pigs, with ventricular fibrillation induced via an implantable defibrillator. Experiment 1 included 2610 annotated heart cycles and used a simple classifier to distinguish compression (only manual) from ROSC signals. Experiment 2 involved 5140 cycles and employed a two- stage classifier: the first stage replicated Experiment 1, while the second further separated compression-only from compression with intrinsic cardiac activity. Two- second spectral signals were extracted, normalized, and artificial neural networks are trained for classifying the signals by using State-of-the-art deep learning models as feature extractors. Grad-CAM, an explainable AI (XAI) method, highlighted key regions which contributed most to the model’s predictions.

Our model achieved mean sensitivity of 98 %, specificity of 97 %, positive predictive value of 97 %, and negative predictive value of 100 %. XAI heatmaps highlighted features important for the model’s predictions.

In a porcine model of cardiac arrest, we demonstrated that deep learning techniques can harness the potential of AI to identify the compressions with intrinsic cardiac activity and ROSC during CPR, achieving highly accurate results.