Learning EKG Diagnostic Models with Hierarchical Class Label Dependencies.

Area of Science:

• Cardiology
• Machine Learning
• Biomedical Informatics

Background:

• Electrocardiogram (EKG/ECG) is a crucial diagnostic tool for assessing cardiac conditions in clinical settings.
• Machine learning (ML) offers potential for automated EKG interpretation, aiding diagnosis and research.
• Current ML approaches often treat EKG diagnostic labels independently, potentially missing complex relationships.

Purpose of the Study:

• To develop and evaluate an ML model that leverages the hierarchical structure of EKG diagnostic labels for improved classification performance.
• To investigate the impact of modeling class-label dependencies on the accuracy of automated EKG interpretation.

Main Methods:

• The proposed ML model transforms EKG signals into a low-dimensional representation.
• A conditional tree-structured Bayesian network (CTBN) is employed to capture hierarchical dependencies among diagnostic labels.
• The model's performance is evaluated on the PTB-XL dataset using multiple classification metrics.

Main Results:

• The CTBN-based model demonstrated improved diagnostic performance compared to models that predict labels independently.
• Modeling hierarchical dependencies between EKG diagnostic classes enhanced classification accuracy.
• The approach showed benefits across various performance metrics for EKG interpretation.

Conclusions:

• Incorporating hierarchical class-label dependencies into ML models enhances EKG diagnostic accuracy.
• The CTBN approach offers a promising method for more sophisticated automated EKG interpretation.
• This work contributes to advancing ML applications in cardiac diagnostics.