Julia Amann, Dennis Vetter, Stig Nikolaj Blomberg, Helle Collatz Christensen, Megan Coffee, Sara Gerke, Thomas K. Gilbert, Thilo Hagendorff, Sune Holm, Michelle Livne, Andy Spezzatti, Inga Strümke, Roberto V. Zicari, Vince Istvan Madai
Explainability for artificial intelligence (AI) in medicine is a hotly debated topic. Our paper presents a review of the key arguments in favor and against explainability for AI-powered Clinical Decision Support System (CDSS) applied to a concrete use case, namely an AI-powered CDSS currently used in the emergency call setting to identify patients with life-threatening cardiac arrest. More specifically, we performed a normative analysis using socio-technical scenarios to provide a nuanced account of the role of explainability for CDSSs for the concrete use case, allowing for abstractions to a more general level. Our analysis focused on three layers: technical considerations, human factors, and the designated system role in decision-making.
Machine learning (ML) powered Artificial intelligence (AI) methods are increasingly applied in the form of Clinical Decision Support Systems (CDSSs) to assist healthcare professionals (HCPs) in predicting patient outcomes. These novel CDSSs have the capacity to propose recommendations based on a plethora of patient data at a much greater speed than HCPs . In doing so, they have the potential to pave the way for personalized treatments, improved patient outcomes, and reduced health care costs.
In laboratory settings, proof-of-concept AI-based CDSSs show promising performance [2,3]. In practice, however, AI-based CDSSs often yield limited improvements [1,4–9]. A possible explanation for this might be that in cases where the AI system’s suggested course of action deviates from established clinical guidelines or medical intuition, it can be difficult to convince HCPs to consider the systems’ recommendations rather than dismissing them a priori.
There is considerable uncertainty about the expected utility and appropriate implementation of explainability in AI-based CDSSs. To advance current theoretical considerations on the topic, we performed a normative analysis using socio-technical scenarios to provide a nuanced account of the role of explainability for CDSSs in a concrete use case. More specifically, we adopted a bottom-up approach where our analysis was informed by a real-world application with the aim to make abstractions for AI-powered CDSS, more broadly. We focused our analysis on three layers: technical considerations, human factors, and the designated system role in decision-making.
In the following, we present the two socio-technical scenarios that outline the implications the foregoing or addition of explainability would have for the use case at hand and what measures could be adopted, respectively, to increase the dispatchers’ trust in the system.
Our findings suggest that omitting explainability will lead to challenges in gaining the users’ trust in the model. Adding explainability could foster such trust. It is, however, challenging to ensure the validity and usefulness of explanations to users of the system. This is the case because explanations must be tailored to the specific use case, that may vary in terms of technical aspects, users and the designated role of the system. In contrast to simple models that often provide a direct form of explanation, in a black-box model as given in our use case, the challenge remains to provide explanations that are supported by robust validation.
We conclude that whether explainability can provide added value to CDSSs depends on several key questions: technical feasibility, the level of validation in case of explainable algorithms, the exact characteristics of the context in which the system is implemented (e.g., the time criticality of the decision making), the designated role in the decision-making process (algorithm-based, -driven, or -determined), and the key user group(s). We deem it likely that the role of explainability cannot be answered definitively at a high, theoretical level. Instead, each system developed for the clinical setting will require an individualized assessment of explainability needs.
Citation: Amann J, Vetter D, Blomberg SN, Christensen HC, Coffee M, Gerke S, et al. (2022) To explain or not to explain?—Artificial intelligence explainability in clinical decision support systems. PLOS Digit Health 1(2): e0000016. https://doi.org/10.1371/journal.pdig.0000016
Editor: Henry Horng-Shing Lu, National Yang Ming Chiao Tung University, TAIWAN
Received: September 28, 2021; Accepted: January 3, 2022; Published: February 17, 2022
Copyright: © 2022 Amann et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: All data are in the manuscript and/or supporting information files.
Funding: JA was supported by funding from the European Union’s Horizon 2020 research and innovation program under grant agreement No. 777107 (PRECISE4Q). DV received funding from the European Union’s Horizon 2020 Research and Innovation Program ‘PERISCOPE: Pan European Response to the ImpactS of COvid-19 and future Pandemics and Epidemics’ under grant agreement no. 101016233, H2020-SC1-PHE-CORONAVIRUS-2020-2-RTD and from the European Union’s Connecting Europe Facility program ‘xAIM: eXplainable Artificial Intelligence for healthcare Management’ under grant agreement no. INEA/CEF/ICT/A2020/2276680, 2020-EU-IA-0098. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: VIM reported receiving personal fees from ai4medicine outside the submitted work. There is no connection, commercial exploitation, transfer or association between the projects of ai4medicine and the results presented in this work.