LDIAED: A lightweight deep learning algorithm implementable on automated external defibrillators

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Automated External Defibrillators (AEDs) are crucial for cardiac arrest resuscitation.
• Accurate differentiation of shockable versus non-shockable Electrocardiogram (ECG) rhythms is critical for AED efficacy.
• Current limitations exist in prompt and precise ECG signal analysis for AEDs.

Purpose of the Study:

• To develop and evaluate a Deep Neural Network (DNN) algorithm for rapid classification of shockable and non-shockable ECG signals.
• To assess the performance of the DNN model in terms of accuracy and latency for real-time AED application.
• To validate the model's applicability on a compact computing platform meeting medical device standards.

Main Methods:

• A Deep Neural Network (DNN) algorithm was designed to analyze 1.4-second ECG signal segments.
• The model was trained using diverse ECG datasets from MIT-BIH, Ventricular Fibrillation Database (VFDB), and Creighton University Database (CUDB).
• The optimized DNN model was deployed on a Raspberry Pi minicomputer for real-time performance testing.

Main Results:

• The DNN algorithm achieved a high accuracy of 99.1% in distinguishing shockable from non-shockable ECG signals.
• The implemented model demonstrated an average latency of 0.845 seconds on the Raspberry Pi, meeting IEC 60601-2-4 standards.
• The frequency-independent nature of the algorithm ensures robustness across various signal conditions.

Conclusions:

• The proposed DNN-based technique offers a highly accurate and prompt method for ECG rhythm classification.
• The successful deployment on a Raspberry Pi indicates the potential for integration into next-generation AED devices.
• This advancement could significantly improve the effectiveness of emergency cardiac resuscitation protocols.