Application of Pre-Trained Deep Learning Models for Clinical ECGs

Area of Science:

• Cardiology
• Medical Informatics
• Artificial Intelligence in Medicine

Background:

• Automatic electrocardiogram (ECG) analysis is a long-standing application of computer-assisted diagnosis (CAD).
• Deep learning (DL) models are emerging as powerful tools for ECG analysis, with some claiming superior performance to human physicians.
• Existing ECG devices offer some level of automatic analysis, necessitating evaluation of new technologies.

Purpose of the Study:

• To assess the added clinical value of a published DL model compared to existing built-in ECG analysis.
• To validate the performance of a DL model in detecting specific arrhythmias like left bundle branch block and atrial fibrillation.

Main Methods:

• A pilot study analyzed 29 12-lead ECGs using a published DL model.
• Results from the DL model were compared against the device's built-in analysis and clinical diagnosis.
• Minor discrepancies in DL model results were noted, attributed to runtime environment differences, but did not impact final classification.

Main Results:

• The DL model successfully reproduced the reported excellent performance in detecting left bundle branch block and atrial fibrillation.
• The DL method and the built-in analysis method demonstrated similar clinical relevance for the selected ECG cases.
• No significant impact on final classification was observed despite minor runtime-related result variations.

Conclusions:

• Deep learning methods show promise for ECG analysis, particularly in research and training settings.
• The clinical utility of DL in routine practice requires further investigation with more extensive and complex cases.
• While DL performance is promising, its added value over current built-in analysis needs comprehensive validation in diverse clinical scenarios.