The prospect of patient harm caused by the decisions made by an artificial intelligence-based clinical tool is something to which current practices of accountability and safety worldwide have not yet adjusted. We focus on two aspects of clinical artificial intelligence used for decision-making: moral accountability for harm to patients; and safety assurance to protect patients against such harm. Artificial intelligence-based tools are challenging the standard clinical practices of assigning blame and assuring safety. Human clinicians and safety engineers have weaker control over the decisions reached by artificial intelligence systems and less knowledge and understanding of precisely how the artificial intelligence systems reach their decisions. We illustrate this analysis by applying it to an example of an artificial intelligence-based system developed for use in the treatment of sepsis. The paper ends with practical suggestions for ways forward to mitigate these concerns. We argue for a need to include artificial intelligence developers and systems safety engineers in our assessments of moral accountability for patient harm. Meanwhile, none of the actors in the model robustly fulfil the traditional conditions of moral accountability for the decisions of an artificial intelligence system. We should therefore update our conceptions of moral accountability in this context. We also need to move from a static to a dynamic model of assurance, accepting that considerations of safety are not fully resolvable during the design of the artificial intelligence system before the system has been deployed.
Established approaches to assuring safety-critical systems and software are difficult to apply to systems employing ML where there is no clear, pre-defined specification against which to assess validity. This problem is exacerbated by the opaque nature of ML where the learnt model is not amenable to human scrutiny. XAI methods have been proposed to tackle this issue by producing human-interpretable representations of ML models which can help users to gain confidence and build trust in the ML system. However, little work explicitly investigates the role of explainability for safety assurance in the context of ML development. This paper identifies ways in which XAI methods can contribute to safety assurance of ML-based systems. It then uses a concrete ML-based clinical decision support system, concerning weaning of patients from mechanical ventilation, to demonstrate how XAI methods can be employed to produce evidence to support safety assurance. The results are also represented in a safety argument to show where, and in what way, XAI methods can contribute to a safety case. Overall, we conclude that XAI methods have a valuable role in safety assurance of ML-based systems in healthcare but that they are not sufficient in themselves to assure safety.