Hunting Bunnies: Comparison of XAI methods for detection of right bundle branch blocks in 12-lead electrocardiograms

Jacqueline Michelle Metsch

Abstract:

Multiple explainable Artificial Intelligence (XAI)
methods have been introduced in the last years to help increase
trust in black-box AI models, especially in the clinical domain.
Due to the variety of available methods it is becoming necessary
to evaluate and compare their performance in revealing what
a model learned. In this work, we tested 15 post-hoc XAI
methods on a real world 12-lead electrocardiogram (ECG)
dataset processed by a deep neural network (DNN) for detecting
Right Bundle Branch Block (RBBB). Only five XAI methods
were able to detect the characteristic two R-peaks (”bunny
ears”) highlighting that the result of XAI analysis is influenced
by the selected XAI method, sample characteristics, and model
structure. The question arises whether XAI methods should be
tested in context of used model and samples before its results are
analyzed to minimize the observed bias.