A review of explainable artificial intelligence for medicine

Large volumes of data have been accumulated in the medical process, mainly including the preventive, diagnostic, and therapeutic processes. Artificial intelligence (AI) systems constructed based on the accumulated data provide a good way to aid clinicians in performing activities more effectively and efficiently. Different from other domains, medicine places significant requirements on both accuracy and explainability due to the strong causality in clinicians’ activities, which induces the focus of this study, explainable AI (XAI) for medicine. From the points of concepts, characteristics, and dimensions, the bases of XAI are introduced. According to the bases, two representative types of existing studies on XAI, including transparent model and post hoc interpretability, and existing studies on the evaluation of XAI are reviewed. How XAI is applied in the preventive, diagnostic, and therapeutic processes is also reviewed. Based on the review of XAI and its application in medicine, transparent models, post hoc interpretability, and XAI evaluation for medicine are discussed to reveal the applicability and insufficiency of existing methods for the preventive, diagnostic, and therapeutic processes. To promote the widespread use of XAI in medicine, future studies on transparent models, post hoc interpretability, and XAI evaluation for medicine are described.