Quantitative Evaluation of xAI Methods for Multivariate Time Series - A Case Study for a CNN-Based MI Detection Model

This paper presents an evaluation framework for xAI methods that is tailored for multivariate time series data. The framework includes three evaluation approaches encompassing a stability analysis, consistency analysis, and truthfulness analysis. The stability analysis investigates the consistency of explanations provided by a single xAI method for similar inputs. In the truthfulness analysis, the meaningfulness of explanations provided by an xAI method is examined. The consistency analysis assesses the similarity of explanations generated by different xAI methods. We demonstrate the application of these evaluation techniques using a medical use case involving electrocardiogram (ECG) data. Specifically, we evaluate the explanations of two popular xAI methods, LRP and SHAP, for a convolutional neural network (CNN) that detects myocardial infarctions (MI). We will show that LRP and SHAP both provide meaningful explanations for this model, with SHAP being slightly more truthful. On the other hand, our stability analysis will reveal that LRP is more stable than SHAP for the investigated use case. Finally, the consistency analysis will allow us to demonstrate that LRP and SHAP partly disagree in explaining the leads and time intervals most relevant for the MI detection model towards its classification.