[Atrial fibrillation diagnosis algorithm based on improved convolutional neural network].

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Atrial fibrillation (AF) is a prevalent arrhythmia linked to increased risks of thrombosis, stroke, and mortality.
• Current screening methods for AF often face challenges with high false-negative rates (FNR), potentially delaying critical treatment.
• Accurate and timely diagnosis of AF is crucial for preventing severe complications.

Purpose of the Study:

• To develop and validate a deep learning model, a convolutional neural network with a low false-negative rate (LFNR-CNN), to improve AF screening.
• To enhance the traditional cross-entropy loss function by incorporating regularization coefficients to penalize false negatives more heavily.
• To reduce the FNR in AF detection without compromising overall diagnostic accuracy.

Main Methods:

• A modified convolutional neural network (CNN) architecture was designed, incorporating an adjusted cross-entropy loss function with regularization coefficients.
• The proposed LFNR-CNN was trained and validated using a large inter-patient clinical database (CD-21077) comprising 21,077 patients.
• The performance of the LFNR-CNN was compared against a standard CNN using traditional cross-entropy loss.

Main Results:

• The LFNR-CNN significantly reduced the FNR from 2.22% to 0.97% compared to the traditional approach.
• Sensitivity (SE) improved from 97.78% to 98.35%, while accuracy (ACC) saw a slight increase from 96.49% to 96.62%.
• The modified loss function effectively minimized false negatives without a substantial impact on overall accuracy.

Conclusions:

• The proposed LFNR-CNN model demonstrates superior performance in detecting atrial fibrillation by substantially lowering the FNR.
• This deep learning approach enhances diagnostic capabilities, reducing the likelihood of missed diagnoses and ensuring timely intervention.
• The developed universal loss function holds potential for improving auxiliary diagnoses in various other clinical applications.