Abstract
Feature importance methods promise to provide a ranking of features according to importance for a given classification task. A wide range of methods exist but their rankings often disagree and they are inherently difficult to evaluate due to a lack of ground truth beyond synthetic datasets. In this work, we put feature importance methods to the test on real-world data in the domain of cardiology, where we try to distinguish three specific pathologies from healthy subjects based on ECG features comparing to features used in cardiologists' decision rules as ground truth. We found that the SHAP and LIME methods and Chi-squared test all worked well together with the native Random forest and Logistic regression feature rankings. Some methods gave inconsistent results, which included the Maximum Relevance Minimum Redundancy and Neighbourhood Component Analysis methods. The permutation-based methods generally performed quite poorly. A surprising result was found in the case of left bundle branch block, where T-wave morphology features were consistently identified as being important for diagnosis, but are not used by clinicians.