A Multi-Center Trained Residual Neural Network for Robust Classification of Atrial High-Rate Episodes in Remotely...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Remote monitoring of pacemakers and defibrillators enhances patient safety but increases clinical workload.
• Reviewing atrial high-rate episodes is time-consuming due to diverse episode types including atrial tachycardia/fibrillation (AT/AF), noise, and far-field oversensing (FFO).

Purpose of the Study:

• To develop and evaluate an Artificial Intelligence (AI) model for automatic review of atrial high-rate episodes in remote pacemaker monitoring.
• To decrease the clinical workload associated with remote monitoring while maintaining high sensitivity for true atrial tachycardia detection.

Main Methods:

• A residual network was trained using center-level fourfold cross-validation.
• Four models were created and combined into an ensemble model by averaging prediction probabilities.
• The ensemble model's performance was evaluated by thresholding predictions at >95% probability.

Main Results:

• Individual models achieved high precision for AT/AF (97.2-99.4%), noise (93.1-97.7%), and FFO (75.4-94.4%).
• High sensitivity for AT/AF was maintained across models (98.9-99.3%).
• The final ensemble model made only two errors (<0.1%) across 3925 reviewed episodes, demonstrating robust performance.

Conclusions:

• AI models can reliably assist in the remote monitoring of pacemakers and defibrillators by automating the review of atrial high-rate episodes.
• The developed AI ensemble model significantly reduces the workload while maintaining high accuracy.
• Future research should focus on AI models for other episode types and clinical validation for widespread adoption.