A systematic approach to enhance the explainability of artificial intelligence in healthcare with application to diagnosis of diabetes

• Published in: Healthcare analytics (New York, N.Y.) Vol. 3; p. 100183
• Main Authors: Wang, Yu-Cheng, Chen, Tin-Chih Toly, Chiu, Min-Chi
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 01.11.2023; Elsevier
• Subjects: Artificial intelligence; Diabetes diagnosis; Explainable artificial intelligence; Healthcare; Local interpretable model-agnostic explanation; Explainable artificial intelligence; Healthcare; Diabetes diagnosis; Artificial intelligence; Local interpretable model-agnostic explanation
• ISSN: 2772-4425; 2772-4425
• DOI: 10.1016/j.health.2023.100183

Explainable artificial intelligence (XAI) tools are used to enhance the applications of existing artificial intelligence (AI) technologies by explaining their execution processes and results. In most past research, XAI tools and techniques are typically applied to only the inference part of the AI application. This study proposes a systematic approach to enhance the explainability of AI applications in healthcare. Several AI applications for type 2 diabetes diagnosis are taken as examples to illustrate the applicability of the proposed methodology. According to experimental results, the XAI tools and technologies in the proposed methodology were more diverse than those in the past research. In addition, an artificial neural network was approximated to a simpler and more intuitive classification and regression tree (CART) using local interpretable model-agnostic explanation (LIME). The extracted rules were used to recommend actions to the users to restore their health. •A systematic approach is proposed to enhance the explainability of artificial intelligence applications for type 2 diabetes diagnosis.•Seven explainable artificial intelligence tools and techniques are applied to various parts of the artificial intelligence application.•The explainable artificial intelligence tools and technologies used are more diverse than those in past research.•A simpler and more intuitive regression tree classification successfully approximates the artificial neural network.