Detecting atrial fibrillation by deep convolutional neural networks.

Area of Science:

• Cardiology
• Artificial Intelligence
• Signal Processing

Background:

• Atrial fibrillation (AF) is the most prevalent cardiac arrhythmia, with incidence rising with age.
• AF significantly increases risks of stroke, morbidity, and mortality.
• Accurate ECG-based AF detection is clinically valuable yet challenging.

Purpose of the Study:

• To propose a novel, reliable, and accurate deep learning method for AF detection.
• To develop deep convolutional neural network models utilizing STFT and SWT for ECG analysis.
• To evaluate and compare the proposed method against existing algorithms.

Main Methods:

• ECG segments were analyzed using Short-Time Fourier Transform (STFT) and Stationary Wavelet Transform (SWT) to create 2-D matrix inputs.
• Two distinct deep convolutional neural network models were developed, one for STFT and one for SWT inputs.
• The method bypasses the need for P or R peak detection and manual feature engineering.

Main Results:

• The proposed method achieved high performance on short ECG segments (as brief as 5s).
• The STFT-based model showed 98.34% sensitivity, 98.24% specificity, and 98.29% accuracy.
• The SWT-based model achieved 98.79% sensitivity, 97.87% specificity, and 98.63% accuracy.

Conclusions:

• The deep learning approach demonstrates high sensitivity, specificity, and accuracy for AF detection.
• This novel method is a valuable tool for clinical AF diagnosis.
• The proposed technique offers an efficient and reliable alternative to traditional AF detection methods.