h1

TY  - THES
AU  - Kempt, Hendrik
TI  - Ethical investigations of AI in medical diagnostics
PB  - Rheinisch-Westfälische Technische Hochschule Aachen
VL  - Dissertation
CY  - Aachen
M1  - RWTH-2023-05683
SP  - 1 Online-Ressource
PY  - 2023
N1  - Veröffentlicht auf dem Publikationsserver der RWTH Aachen University
N1  - Dissertation, Rheinisch-Westfälische Technische Hochschule Aachen, 2023
AB  - Since the advent of machine learning as an artificial intelligence paradigm, opportunities for automating clinical processes and decision-making have emerged that were previously considered unattainable. This dissertation aims to develop ethical analyses of the issues surrounding the use of "medical AI" for various diagnostic purposes. To motivate and contextualize the articles, a brief overview of the developments of artificial intelligence in medical diagnostics and the related ethical debate is first presented. For this endeavor, three different levels are presented against which diagnostic AI can be analyzed in terms of its intent and function. The first level concerns the role of AI as a decision support system. In this role, AI is primarily understood according to the assistance of physicians and patients in the decision-making process. The requirements for AI in this role are thus to compensate for human limitations. At the same time, such assistance can not only support decisions, but also make new decisions possible in the first place. The second level assumes that AI can replace human decisions, i.e. those of both physicians and patients. In this view, AI applications for diagnostic purposes become successful when they exceed the benchmarks of human decision-making. As a third level, the influence of AI on medical practice is examined. Here, the role of AI as an assistant or replacement is less important than the consequences that the use of such AI may have for other medical processes. This concerns, for example, norms of the doctor-patient relationship. This level allows for an analysis of the impact of AI on medical norms that goes beyond the analysis of the technology. The first paper discusses several of these conceptual and normative issues and proposes a solution to a specific use case of AI for the provision of second medical opinions. To this end, a previously unavailable taxonomy of types of clinical second opinions is first undertaken. With an overview of the concepts of responsibility, explainability, and peer disagreements, this paper points to a pragmatic yet normative solution that allows AI-based decision support systems to take on as many tasks in diagnostic processes as possible: In the case of a second opinion of an AI confirming the initial diagnosis, this can be considered sufficient evidence for a diagnosis; should such a second opinion deviate from the initial diagnosis of the treating physician, this must be considered as an indicator that a further, human assessment of that case is necessary. This norm is referred to here as the "rule of disagreement". The second paper builds on these findings and addresses the question of how decision support systems can be disagreed with in clinical contexts. Physicians who come to different diagnostic conclusions than an AI they use as such a decision support system face a double challenge: on the one hand, the AI's diagnostic proposal is evidence that needs to be taken into account; on the other hand, the decision for a diagnosis is in the hands of physicians, and a rejection of the evidence produced by the AI requires further, alternative justifications that are not readily available. To this end, we first distinguish the forms of conflict between physicians and technology: malfunctions, mistakes, and disagreements. Starting from the notion of Meaningful Human Control (MHC) as a normative standard for assessing the ethical permissibility of cooperative systems in clinical processes, the notion of meaningful disagreement is introduced and elaborated. In the third part, the notion of explainability of AI-generated diagnoses is normatively examined. To this end, the various goods that explainability of AI diagnoses may entail are first described. Through the conceptual distinction between äbsolute" and "relative" explainability, a normative distinction is then also introduced to examine the standards against which the explanations of an AI diagnosis are measured. Insofar as such explanations are understood in comparison to human standards ("relative") of diagnoses rather than absolute, higher claims for AI diagnoses cannot be held. Last, it is shown that the goods of explainability presented at the beginning can be replaced without loss by recourse to appropriate risk assessment through certification ("certifiability") and reference to the physician's expertise in translating the AI-based diagnosis to the patient’s understanding (ïnterpretability"). The fourth paper continues the distinctions made in the third regarding the relative and absolute explainability of AI diagnoses and transfers them to the context of global standardizations of AI systems and therefore additionally distinguishes between local and global standards of explainability. High standards of explainable diagnoses in countries with high medical standards create a risk that AIs developed in these countries may not be approved globally due to insufficient standards of explainability; to avoid regions with a lack of experts in explaining diagnoses being unable to use these AIs, it is suggested that the decision to set standards should be left to the communities themselves. The fifth part examines the extent to which AI diagnostic tools should be used as placebos. For this purpose, the debate about placebos in clinical contexts is first examined and a distinction is made between seven different forms of placebo effect eliciting tools. An analysis of doctor-patient relationships as an intimate relationship of trust and the resulting "normative entanglement" establishes the permissibility of placebos. The next step is to examine the skill of AI applications as placebos. Since an AI cannot replace the normative entanglement of doctors and patients, its admissibility is examined solely in the context of trusting doctor-patient treatments. In the sixth part, the question is answered whether an AI should make judgments of medical interventions as "medically necessary". To this end, we first introduce the distinction between medical necessity as a judgment about the instrumental necessity of an intervention and as a judgment about the socially justifiable minimum of medical services (ßocial medical necessity"). By focusing on the first distinction, questions of judgment are raised by AIs. Since these judgments always include judgments about the goals of an intervention, which are socially and morally agreed upon, a skeptical position of AI as a judge of interventions as medically necessary is taken. This is followed by a discussion of both the findings of these papers and the prospect of further research questions in this context. The approach of assessing the uses and influences of AI according to the three levels allows for a problem-oriented, pragmatic view of the use of AI, but also allows for the setting of clear boundaries and the introduction of sharp ethical distinctions. Through these distinctions and the theses connected to them, further issues also emerge, such as the question of democratizing standards of explicability, global AI regulations and growing inequalities in healthcare, new attributions of responsibility in physician-AI collaborations, and changes in clinical work and care environments.
LB  - PUB:(DE-HGF)11
DO  - DOI:10.18154/RWTH-2023-05683
UR  - https://publications.rwth-aachen.de/record/959574
ER  -