ECG Feature Importance Rankings: Cardiologists vs. Algorithms.

Area of Science:

• Cardiology
• Machine Learning
• Biomedical Informatics

Background:

• Feature importance methods aim to rank features for classification tasks.
• Existing methods often yield disagreeing rankings and are hard to validate on real-world data.
• Cardiology, specifically Electrocardiogram (ECG) analysis, presents a domain with established clinical decision rules for ground truth comparison.

Purpose of the Study:

• To evaluate the performance of various feature importance methods on real-world cardiology data.
• To compare the rankings from different methods against established clinical decision rules for ECG interpretation.
• To identify reliable feature importance techniques for clinical applications in cardiology.

Main Methods:

• Applied multiple feature importance techniques (SHAP, LIME, Chi-squared, MRMR, NCA, permutation-based) to ECG data for classifying cardiac pathologies.
• Utilized cardiologist's decision rules based on ECG features as the ground truth for comparison.
• Compared feature rankings from machine learning models (Random Forest, Logistic Regression) with those from feature importance methods.

Main Results:

• SHAP, LIME, and Chi-squared tests showed good agreement with Random Forest and Logistic Regression rankings.
• Maximum Relevance Minimum Redundancy (MRMR) and Neighbourhood Component Analysis (NCA) produced inconsistent results.
• Permutation-based methods generally performed poorly in this cardiology context.
• T-wave morphology features were identified as important for left bundle branch block diagnosis, despite not being used by clinicians.

Conclusions:

• SHAP, LIME, and Chi-squared tests are promising feature importance methods for cardiology applications when combined with Random Forest and Logistic Regression.
• Certain feature importance methods demonstrate unreliability for real-world clinical data.
• Clinical insights from feature importance methods may uncover novel diagnostic indicators, such as T-wave morphology in left bundle branch block.