Mixed convolutional and long short-term memory network for the detection of lethal ventricular arrhythmia

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence in Medicine

Background:

• Early defibrillation is critical for out-of-hospital cardiac arrest (OHCA) survival.
• Automated external defibrillators (AEDs) rely on accurate ventricular fibrillation (VF) detection for shock decisions.
• Deep learning, particularly 1D Convolutional Neural Networks (CNNs), shows promise for VF detection.

Purpose of the Study:

• To develop and evaluate a novel deep learning architecture for enhanced VF detection.
• To compare the proposed architecture against existing methods for VF detection in AED algorithms.
• To assess the algorithm's performance on both public Holter data and real-world OHCA patient data.

Main Methods:

• A deep learning architecture combining 1D-CNN layers and a Long Short-Term Memory (LSTM) network was developed.
• Two datasets were utilized: public Holter recordings and OHCA patient data.
• Performance was evaluated using balanced accuracy (BAC), sensitivity (Se), and specificity (Sp), compared against 1D-CNN and Support Vector Machine (SVM) models.

Main Results:

• The proposed CNN-LSTM architecture achieved high performance: 99.3% BAC on public data and 98.0% BAC on OHCA data.
• It outperformed existing methods by at least 0.3% (public data) and 2.2% (OHCA data) in BAC.
• The algorithm met American Heart Association criteria (95% Sp, 90% Se) for ECG segments as short as 3 seconds.

Conclusions:

• The developed deep learning architecture demonstrates superior accuracy for VF detection, especially in OHCA scenarios.
• This algorithm enables rapid and precise shock/no-shock decisions for AEDs.
• It represents a significant advancement in automated cardiac arrest management technology.