Fully Convolutional Deep Neural Networks with Optimized Hyperparameters for Detection of Shockable and Non-Shockable...

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Cardiology

Background:

• Deep neural networks (DNNs) offer high diagnostic accuracy for electrocardiogram (ECG) analysis but require extensive training and optimization.
• Limited research exists on optimizing DNNs for shock advisory systems using large out-of-hospital cardiac arrest (OHCA) ECG databases.

Purpose of the Study:

• To optimize hyperparameters of deep convolutional neural networks (CNNs) for distinguishing shockable (Sh) and non-shockable (NSh) cardiac rhythms.
• To validate the optimized CNNs for short (2-10 seconds) ECG analysis durations in OHCA patients.

Main Methods:

• Trained and validated an end-to-end deep CNN architecture on a large dataset of OHCA ECGs (720+3170 training, 739+5921 validation).
• Employed random search to optimize hyperparameters (number of layers, filters, kernel size) and selected models based on maximal balanced accuracy (BAC).
• Validated the best-performing model across various analysis durations (2-10 seconds).

Main Results:

• The optimal CNN models, particularly those with over three convolutional layers, achieved high performance (BAC 99.31-99.5%).
• The best model ({N=5, Fi={20,15,15,10,5}, Ki={10,10,10,10,10}}) demonstrated maximal validation performance at 5-second analysis (BAC=99.5%, Se=99.6%, Sp=99.4%).
• A minimal performance drop (<2% points) was observed for very short 2-second analysis (BAC=98.2%).

Conclusions:

• Optimized DNNs, specifically CNNs, can significantly enhance the performance of shock advisory systems in detecting cardiac arrhythmias.
• The developed models can shorten ECG analysis duration, aligning with resuscitation guidelines and minimizing hands-off intervals during cardiac arrest events.