Towards Trustworthy AI in Cardiology: A Comparative Analysis of Explainable AI Methods for Electrocardiogram Interpretation

Deep learning models have successfully been applied for medical diagnoses. However, such models are perceived as opaque black boxes and their decisions are not comprehensible for human users. This intransparency and the potential resulting lack of trust is still an obstacle for a wider use of deep learning models in medicine. Explainable Artificial Intelligence (XAI) is expected to become an important factor for trust, transparency, and accountability of AI models. We present a comparison of state-of-the-art XAI methods in terms of their ability to recognize medically relevant patterns in electrocardiogram (ECG) data according to criteria defined in established medical guidelines. We trained separate convolutional neural networks for four common cardiac pathologies and applied state-of-the-art XAI methods to compare the input features identified as important by these methods. Further analyses assessed the patterns identified as relevant in terms of their distribution across leads and their overall overlap with guideline criteria. Across most XAI methods, we observed high conformance between the relevance assignment to the patterns and ECG leads listed in the referenced guidelines. However, it turned out that the explanations varied significantly depending on the choice of the XAI method. This study points out strengths and weaknesses of the XAI methods under examination. Overall, our experiments show that XAI can increase the transparency of deep learning models in showing how patterns relevant for the algorithm relate to human decision making. This may serve as an essential aspect for the implementation of trustworthy AI in cardiology.