The effect of confounding data features on a deep learning algorithm to predict complete coronary occlusion in a...

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Deep learning (DL) is increasingly utilized for automated electrocardiogram (ECG) analysis.
• Early detection of coronary artery occlusion is critical for patient outcomes.
• Evaluating DL algorithms against expert cardiologists is essential for clinical adoption.

Purpose of the Study:

• To assess the feasibility of using a DL algorithm for early detection of induced coronary artery occlusion.
• To compare the DL model's performance against experienced cardiologists and STEMI criteria.
• To investigate the potential of DL in identifying acute coronary artery occlusion from ECG data.

Main Methods:

• A retrospective observational study employed a deep convolutional neural network trained on the STAFF III database.
• The DL model classified ECG samples as either acute coronary artery occlusion or no occlusion.
• Cross-validation was used to evaluate performance, with two iterations using different non-occluded sample sources.

Main Results:

• The first iteration achieved a high F1 score (0.814), surpassing cardiologists and STEMI criteria.
• The second iteration yielded a significantly lower F1 score (0.533), comparable to random chance.
• Transfer learning was used in the second iteration, but the dataset size limited performance.

Conclusions:

• Data leakage in the first iteration led to artificially inflated performance results.
• The study underscores the critical risk of DL models producing spurious results due to data leaks.
• Rigorous validation is paramount to ensure the reliability of DL algorithms in clinical settings.