Deep Neural Networks for ECG-Based Pulse Detection during Out-of-Hospital Cardiac Arrest

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence in Medicine

Background:

• Automatic pulse detection is crucial for recognizing out-of-hospital cardiac arrest (OHCA) and return of spontaneous circulation.
• The electrocardiogram (ECG) is the sole universally available signal in defibrillators for pulse detection.
• Distinguishing between pulseless electrical activity (PEA) and pulse-generating rhythm (PR) is vital for OHCA management.

Purpose of the Study:

• To develop and evaluate deep neural network (DNN) architectures for automatic pulse detection using ECG segments.
• To classify cardiac rhythms as PEA or PR using short ECG recordings.
• To compare the performance of proposed DNNs against existing state-of-the-art methods.

Main Methods:

• Two DNN architectures were designed: a fully convolutional neural network and a convolutional network with an added recurrent layer.
• Bayesian optimization was employed for hyperparameter tuning of both DNN models.
• A dataset of 3914 5-second ECG segments from 279 OHCA episodes was used, partitioned into training (80%) and testing (20%) sets.

Main Results:

• The first DNN achieved a sensitivity (Se) of 94.1%, specificity (Sp) of 92.9%, and balanced accuracy (BAC) of 93.5%.
• The second DNN architecture demonstrated improved performance with Se 95.5%, Sp 91.6%, and BAC 93.5%.
• Both DNN models outperformed current state-of-the-art PR/PEA discrimination algorithms, showing over 1.5 points improvement in BAC.

Conclusions:

• Deep neural networks offer a promising approach for accurate pulse detection from ECG in OHCA scenarios.
• The proposed DNN architectures effectively discriminate between PEA and PR rhythms.
• These findings suggest potential for improved automated decision support in OHCA management using AI-powered ECG analysis.