Rhythm Analysis during Cardiopulmonary Resuscitation Using Convolutional Neural Networks.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Cardiopulmonary resuscitation (CPR) chest compressions create ECG artifacts, potentially leading to incorrect defibrillator rhythm classification.
• Accurate rhythm analysis during CPR is crucial for effective defibrillation and improved patient survival rates.

Purpose of the Study:

• To design and evaluate a novel algorithm utilizing convolutional neural networks (CNNs) for reliable shock/no-shock decisions during CPR.
• To overcome the challenge of ECG artifacts induced by chest compressions.

Main Methods:

• A dataset of 3319 ECG segments (9s each) from CPR, including 586 shockable and 2733 non-shockable rhythms, was analyzed.
• Recursive Least Squares (RLS) filtering removed chest compression artifacts.
• A CNN classifier with three convolutional blocks and two fully connected layers was employed for classification.
• A 5-fold cross-validation, repeated 100 times, was used for robust performance evaluation and comparison against a baseline model.

Main Results:

• The proposed CNN algorithm achieved high performance metrics: median sensitivity of 95.8%, specificity of 96.1%, accuracy of 96.1%, and balanced accuracy of 96.0%.
• The algorithm demonstrated a slight improvement in accuracy (0.6 points) over the best-performing baseline model using handcrafted features and a random forest classifier.
• The deep learning approach proved effective in providing reliable cardiac rhythm diagnosis without interrupting chest compressions.

Conclusions:

• Deep learning methods, specifically CNNs, show significant potential for accurate cardiac rhythm classification during CPR.
• The developed algorithm offers a reliable solution for automated shock/no-shock decisions, even in the presence of significant ECG artifacts.
• This approach may enhance the effectiveness of defibrillation therapy in critical care settings.