Over-fitting suppression training strategies for deep learning-based atrial fibrillation detection.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Signal Processing

Background:

• Deep learning models for atrial fibrillation (AF) detection in electrocardiogram (ECG) signals often suffer from overfitting, leading to reduced accuracy on independent datasets.
• This overfitting issue is particularly pronounced when detecting AF from dynamic ECG recordings.

Purpose of the Study:

• To explore and validate two novel training strategies to mitigate the overfitting problem in deep learning-based AF detection.
• To enhance the robustness and accuracy of AF detection models using ECG data.

Main Methods:

• Implemented Fast Fourier Transform (FFT) and Hanning-window filtering to reduce individual differences in ECG signals.
• Utilized wearable ECG data from 29 arrhythmia patients (collected over 24h) for model training to improve robustness.
• Developed and tested a Long Short-Term Memory (LSTM) and Convolution Neural Network (CNN) hybrid model.

Main Results:

• The proposed model achieved high accuracy rates: 96.23% on independent wearable ECG data, 95.44% on the MIT-BIH AF database, and 95.28% on the PhysioNet Challenge 2017 database.
• Models trained with the proposed strategies showed only a 2% accuracy decrease on independent sets, compared to a 15% decrease for models trained without these strategies.
• The strategies effectively enhanced the detection accuracy and robustness of deep learning networks for AF detection.

Conclusions:

• The explored training strategies, including signal processing and wearable data utilization, are effective in developing robust AF detectors.
• These methods significantly improve the generalization capability and detection accuracy of deep learning models for AF in ECG signals.