Adoption of AI-driven Clinical Decision Support Systems: A Checklist for Healthcare Providers Based on a Narrative Review of AI Evaluation Resources and Expert Interviews

Abstract

Background: Artificial intelligence (AI) has the potential to support healthcare professionals and improve patients’ outcomes. AI-based clinical decision support systems (CDSSs) are reported to be particularly promising. To ensure the suitability of an AI-CDSS and prevent negative impacts, healthcare providers should ask the ‘right’ questions before adoption. However, there is yet no evaluation tool for AI-CDSS adoption publicly available. This thesis aimed to (1) identify guidelines and evaluation tools applicable to the adoption of CDSS and AI, and (2) synthesise AI-CDSS adoption considerations in a checklist for healthcare providers. Methods: Trustworthy AI evaluation tools were previously identified in a scoping review by the author and colleagues. Guidelines and evaluation tools for other pre-identified categories of AI-CDSS adoption considerations were searched in PubMed, Scopus, and Google. Additional data was collected through four semi-structured interviews with experts who have backgrounds in medicine, bioethics and law, and the social science of the internet. The interviews were analysed using thematic analysis, while items from each literature source were categorised to summarise and structure AI-CDSS adoption considerations. Results: A total of 76 literature sources, published between 2011 and 2025 and originating mainly from developed countries, were included. The majority of these sources focused on trustworthy AI or AI maturity, though guidance and evaluation tools related to other adoption categories were also identified. Their items were synthesised into a list of 227 AI-CDSS adoption questions covering the following categories: (1) regulatory and legal compliance, (2) utility, (3) trustworthy AI, (4) economic aspects, (5) usability, (6) workflow integration, (7) AI maturity, and (8) vendor reliability, support, and agreements. The expert interviews verified considerations covered by the list and helped to identify the most relevant ones. They also provided guidance on the development of an AI-CDSS adoption checklist with 20 questions. Conclusions: The checklist integrates findings from 76 literature sources and four expert interviews. It can support both the decision whether an AI-CDSS should be adopted and the deployment of a system. While feedback on the checklist has been received from three experts and incorporated, a Delphi study involving a larger number of experts from diverse disciplines would enhance its usefulness. Furthermore, the checklist’s practicality needs to be tested in the real-world, and it should be updated as the use of AI in healthcare continues to evolve.