The healthcare sector has become a key area for the development and application of new technology and, not least, Artificial Intelligence (AI). New reports are constantly being published about how this algorithm-based technology supports or performs various medical tasks. These illustrates the rapid development of AI that is taking place within healthcare and how algorithms are increasingly involved in systems and medical devices designed to support medical decision-making.  The digital revolution and the advancement of AI technologies represent a step change in the way healthcare may be delivered, medical services coordinated and well-being supported. It could allow for easier and faster communication, earlier and more accurate diagnosing and better healthcare at lower costs. However, systems and devices relying on AI differs significantly from other, traditional, medical devices. AI algorithms are – by nature – complex and partly unpredictable. Additionally, varying levels of opacity has made it hard, sometimes impossible, to interpret and explain recommendations or decisions made by or with support from algorithmic decision systems. These characteristics of AI technology raise important technological, practical, ethical and regulatory issues. The objective of this thesis is to analyse the relationship between the EU regulation on medical devices (MDR) and algorithmic decision systems (ADS) used within the medical domain. The principal question is whether the MDR is enough to guarantee safe and robust ADS within the European healthcare sector or if complementary (or completely different) regulation is necessary. In essence, it will be argued that (i) while ADS are heavily reliant on the quality and representativeness of underlying datasets, there are no requirements with regard to the quality or composition of these datasets in the MDR, (ii) while it is believed that ADS will lead to historically unprecedented changes in healthcare , the regulation lacks guidance on how to manage novel risks and hazards, unique to ADS, and that (iii) as increasingly autonomous systems continue to challenge the existing perceptions of how safety and performance is best maintained, new mechanisms (for transparency, human control and accountability) must be incorporated in the systems. It will also be found that the ability of ADS to change after market certification, will eventually necessitate radical changes in the current regulation and a new regulatory paradigm might be needed.

Cette thèse s'inscrit dans le cadre de la théorie de la décision algorithmique, qui est une discipline au croisement de la théorie de la décision, la recherche opérationnelle et l'intelligence artificielle. Dans cette thèse, nous étudions l'utilisation de plusieurs modèles décisionnels pour résoudre des problèmes de décision séquentielle dans l'incertain, d'optimisation robuste, et d'optimisation multi-agents équitable. Pour résoudre efficacement ces problèmes, nous utilisons des méthodes de type maître-esclaves, dites à base d'oracles dans la thèse. Ces méthodes permettent de résoudre des problèmes de grande taille en procédant de manière incrémentale. Une attention particulière est portée au modèle de l'espérance d'utilité antisymétrique et bilinéaire, au modèle de l'espérance d'utilité pondérée et à leurs pendants en décision multicritère. L'intérêt de ces modèles est multiple. En effet, ils étendent les modèles standards (e.g., modèle de l'espérance d'utilité) et permettent de représenter un spectre étendu de préférences tout en conservant leurs bonnes propriétés théoriques et algorithmiques. La thèse apporte des réponses sur des aspects théoriques (e.g., résultats de complexité algorithmique) et sur des aspects opérationnels (e.g., conception de méthodes de résolution efficaces) aux problèmes soulevés par l'emploi de ces critères dans les contextes susmentionnés. / This thesis falls within the area of algorithmic decision theory, which is at the crossroads between decision theory, operational research and artificial intelligence. In this thesis, we study several decision models to solve problems in different domains: sequential decision problems under risk, robust optimization problems, and fair multi-agent optimization problems. To solve these problems efficiently, we use master-slave algorithms which solve the problem through an incremental process. These procedures, referred to as oracle methods in the thesis, make it possible to solve problems of large size. A particular attention is given to the skew-symmetric bilinear utility model, the weighted expected utility model and their counterparts in multicriteria decision making. These models are interesting at several respects. They extend the standard models (e.g., the expected utility model) and allow to represent a broader class of preferences while retaining their good theoretical and algorithmic properties. The thesis focuses both on theoretic (e.g., complexity results) and operational (e.g., design of practically efficient solution methods) aspects of the problems raised by the use of these criteria in the domains aforementioned.

This thesis develops a new abstraction for solving problems in decision automation. Decision automation is the process of creating algorithms which use data to make decisions without the need for human intervention. In this abstraction, four key ideas/problems are highlighted which must be considered when solving any decision problem. These four problems are the decision problem, the learning problem, the model reduction problem, and the verification problem. One of the benefits of this abstraction is that a wide range of decision problems from many different areas can be broken down into these four “key” sub-problems. By focusing on these key sub-problems and the interactions between them, one can systematically arrive at a solution to the original problem. Three new learning platforms have been developed in the areas of portfolio optimization, business intelligence, and automated water management in order to demonstrate how this abstraction can be applied to three different types of problems. For the automated water management platform a full solution to the problem is developed using this abstraction. This yields an automated decision process which decides how much water to release from the Piute Reservoir into the Sevier River during an irrigation season. Another motivation for developing these learning platforms is that they can be used to introduce students of all disciplines to automated decision making.

Traitement des mégadonnées, surveillance, prédictions comportementales ou aides à la décision, les avantages techniques et commerciaux attribués à l'intelligence artificielle emportent l’engouement et l’adhésion des acteurs économiques privés. Ayant vocation à reproduire les facultés cognitives de l’être humain, l’intelligence artificielle s’immisce ainsi progressivement dans nos activités, nos usages et plus largement, dans nos vies. Pourtant, les défauts de la technologie inquiètent. Utilisée à des fins de reconnaissance faciale, de profilage publicitaire ou encore de recrutement, les biais de l’intelligence artificielle représentent des risques de discriminations pour les personnes qui interagissent avec cette technologie. Or dans un secteur aussi sensible que le recrutement, un tel risque représente un enjeu aussi bien pour les candidats, exposés à une violation de leur droit fondamental à l’égalité, que pour les employeurs qui, eux, s’exposeraient à des sanctions juridiques. En l’absence d’un cadre juridique spécifique à l’intelligence artificielle, la question se pose donc de savoir si notre droit permet l’appréhension de ces formes de discriminations à l’embauche. Le propos de ce mémoire consistera donc à proposer des réponses à cette interrogation en trois temps : l’étude du cadre légal applicable, la gestion du risque de biais discriminatoire et l’enjeu de l’accès à la justice des candidats lésés. / Big data processing, surveillance, behavioral predictions or decision aids, the technical and commercial advantages attributed to artificial intelligence have won the enthusiasm and support of private economic players. Designed to reproduce the cognitive abilities of human beings, artificial intelligence is gradually interfering in our activities, our practices and more widely in our lives. Nevertheless, the flaws of the technology are concerning. Used for facial recognition, advertising profiling or recruitment purposes, the artificial intelligence biases are a risk of discrimination against people who interact with this technology. In a sector as sensitive as recruitment, such a risk constitutes a challenge both for candidates, exposed to a violation of their fundamental right to equality, and for employers who would be exposed to legal sanctions. Without a specific legal framework for artificial intelligence, the question therefore arises as to whether our law allows for the apprehension of these forms of discrimination in hiring. The purpose of this thesis will consist in proposing answers to this question in three stages : the study of the applicable legal framework, the challenges of managing the risk of discriminatory bias, and access to justice for aggrieved candidates.