REMEDIS -Interdisciplinary Considerations and Consequences related to the Development of an AI-supported Medication Dispenser

This thesis presents an interdisciplinary analysis of an engineering problem that stems from developing a particular device, an automated AI-assisted medication dispenser, REMEDIS. The nature of the mechanical engineering solution employed during the course of development to dispense pills brought forward the question about the acceptable level of mistakes regarding a technology, towards which the expectation is to work most of the time but not necessarily all the time. The thesis discusses how and why we created an AI-based solution to verify the pills before they get dispensed to the user and what considerations lead to the unique system design that mitigates the fact that both the engineering solutions and the AI algorithm can only be evaluated in statistical terms. This inherent risk of the technologies deployed in the device prompted theoretical questions we analysed from various angles. The thesis seeks to bring together considerations on how safe our product should be, given the conceptual structure of the problem it tries to mitigate, that of the regulatory environment surrounding it, and of how we can calculate the probability of risks and how to determine whether the risks are not greater than acceptable. First, it reviews the medication adherence literature to understand the adherence characteristics of the potential users of the dispenser seeking to assign quantitative metrics to their successful medication execution rate. As REMEDIS provides an automated dispensing solution that selects the pills for the user, their successful intake ratio serves as a benchmark level that the device must surpass. The thesis also examines whether a numerical upper threshold can be established, a medication-specific adherence rate above which mistakes made by humans or machines have no medical consequences. The thesis reviews the main interpretations of probability theory to determine in what sense risks can be calculated and how compliance with the risk assessment requirements of the legislation can be understood. It addresses the issues various interpretations face when they are evaluated in the context of potential models to determine probabilities in statutory regulations. The thesis provides a detailed analysis of how to select reference classes, particularly for REMEDIS, and how to compile the correct sample so that the statistical results could serve as an adequate predictive model for the unintended consequences of REMEDIS. Seeking answers to product safety requirements and how to establish an acceptable error rate from a legal perspective, the thesis identifies the relevant EU legislation and analyses the texts from the perspective of the research question. In conclusion, the thesis provides an outlook on problems that were raised through the analysis of REMEDIS but are not unique to them. It also discusses the findings in the complex multi-stakeholder environment where the device is intended to be deployed. As a result, the thesis aims to bring to the surface conflicts that otherwise would have remained invisible if boundaries of scientific disciplines were not crossed and the analysis had remained at a sufficiently abstract level, where the applicability of notions, theoretical concepts and regulative texts would not have been put to the test, as in the case of actual product development, where all these aspects influence the entire technical and system design, from major decisions in mechanical engineering to every detail of its software architecture.