12-Lead ECG interpretation by database comparison.

Area of Science:

• Cardiology
• Medical Informatics
• Artificial Intelligence in Medicine

Background:

• Automated electrocardiogram (ECG) interpretation often relies on rule-based expert systems, which can lead to disagreements among experts.
• An alternative approach involves basing automated interpretations on physician-interpreted, similar ECGs, indirectly incorporating varied expert criteria.

Purpose of the Study:

• To develop and evaluate an ECG interpretation algorithm that utilizes the similarity between ECGs to automate analysis.
• To assess the performance of this novel algorithm in classifying various cardiac conditions.

Main Methods:

• A large database of approximately 146,000 12-lead ECGs was utilized.
• The algorithm employed an ECG similarity search and a method to estimate interpretation from similar ECGs.
• Receiver Operating Characteristic (ROC) analysis, including sensitivity, specificity, and area under the curve (AUC), was used for performance testing.

Main Results:

• The algorithm demonstrated high performance across various ECG categories.
• Left Bundle Branch Block (LBBB) achieved the highest AUC of 0.981.
• Other conditions like Right Bundle Branch Block (RBBB), Left Anterior Fascicular Block (LAFB), and Left Ventricular Hypertrophy (LVH) also showed excellent AUC values (≥0.95).

Conclusions:

• ECG interpretation based on similarity analysis provides adequate performance on unselected ECG data.
• While not a replacement for current rule-based systems, this algorithm can serve as a valuable adjunct recommender for ECG interpretation.