A Comparison of Global Explanations Given on Electronic Health Records

Explainable Artificial Intelligence (XAI) is a cutting-edge research area within Artificial Intelligence (AI), playing a crucial role in enhancing predictions in the medical field. However, the formal evaluation of XAI methods remains insufficient. In this study, we delve into state-of-the-art XAI techniques, namely SHapley Additive Predictions (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME), to provide comprehensive explanations using factual and counterfactual data generated by Diverse Counterfactual Explanations (DiCE). To compare these explanation methods effectively, we propose a novel data generative method. Our experiments encompass both factual and generative counterfactual data, enabling us to measure the level of consistency across various XAI methods. We introduce Attribution Space, a new metric for comparing feature attribution methods, alongside well-known measures like the Jaccard similarity index and the Pearson correlation coefficient. Our findings demonstrate that SHAP exhibits greater variability when explaining both factual and counterfactual datasets. Although LIME and SHAP display a strong correlation in certain scenarios, there is seldom agreement on the most significant features returned and the attribution space. This highlights the necessity of employing multiple XAI methods when analyzing Electronic Health Records.