Thèse Ai-Driven Digital Twins For Autonomous And Coordinated Decision-Making In Hospital Pharmaceutical Networks H/F - Doctorat.Gouv.Fr

Établissement : IMT Mines Albi École doctorale : SYSTEMES Laboratoire de recherche : CGI - Centre de Génie Industriel Direction de la thèse : Safa LAYEB ORCID 0000000325367872 Début de la thèse : 2027-09-01 Date limite de candidature : 2026-11-23T23:59:59 Hospital pharmaceutical networks are a critical, and chronically under-optimised, component of healthcare systems. Drug shortages, pharmaceutical waste, near-expiry stock, and misallocation of medicines across hospital sites generate significant clinical risks, financial costs, and environmental harm. Although these networks generate large volumes of operational data (dispensing, stock movements, orders, traceability), this data remains largely unexploited for real-time decision support: replenishment and allocation decisions still rely on static rules or individual judgment.

This thesis aims to develop an autonomous, data-driven decision intelligence system for interconnected hospital pharmaceutical networks, where a central pharmacy supplies several hospital sites with heterogeneous demand profiles. The project is organised around three complementary axes, progressing from understanding the real network to autonomous decision-making.

The first axis builds a process-aware, self-evolving digital twin of the pharmaceutical network from real event logs, relying on Object-Centric Process Mining. The second axis enriches this twin with an AI intelligence layer combining demand prediction, early anomaly detection, and 'What if?' scenario simulation. The third axis constitutes the decision layer: replenishment and allocation policies are learned through reinforcement learning (RL), with a multi-agent extension (MARL) coordinating decisions between the central pharmacy and hospital sites while preserving their local autonomy.

The project relies on a structured partnership combining CHU de Toulouse as the operational testbed, the Centre Génie Industriel at IMT Mines Albi as the host laboratory, and international co-supervision with the LR-OASIS Laboratory at the École Nationale d'Ingénieurs de Tunis (ENIT). This Franco-Tunisian dimension builds on a scientific collaboration already active for several years between the supervisory teams.

This work will contribute to improving care quality, patient safety, and the sustainability of health systems, positioned at the intersection of industrial engineering, artificial intelligence, and hospital management - in full alignment with the 'Understanding and promoting health and well-being' axis of the BEST doctoral programme.
Hospital pharmaceutical networks generate large volumes of data (dispensing, stock, orders, traceability) that remain largely unexploited for real-time decision-making. In an interconnected network where a central pharmacy supplies several hospitals with their own constraints, a decision made at one site constrains the options available elsewhere, while the central pharmacy typically lacks a continuously updated, faithful picture of the network's real state. This gap generates stock shortages, pharmaceutical waste, avoidable emergency transfers, and suboptimal resource use.

Recent literature underscores the urgency of this challenge: pharmaceutical waste is identified as a major economic and environmental issue, the WHO published global guidelines on pharmaceutical waste management in 2025, and a recent scoping review of healthcare digital twins found that only 19% of reviewed studies had been tested in a real environment, with continuous twin updating remaining a largely underdeveloped capability. Moreover, existing work treats digital twins, process mining, and reinforcement learning for pharmaceutical inventory management separately, without integrating them into a single system continuously updated from real data. This thesis aims to bridge that gap.
The project aims to develop an intelligent decision-support system for interconnected hospital pharmaceutical networks, structured around three complementary scientific objectives:
1. Build a process-aware, self-evolving digital twin of the pharmaceutical network, capable of automatically detecting deviations between the model and real operational conditions.
2. Enrich this twin with an AI intelligence layer providing demand prediction, early anomaly detection, and disruption scenario simulation.
3. Develop an autonomous decision layer, based on centralised and then multi-agent reinforcement learning, capable of proposing coordinated replenishment and allocation policies across multiple hospital sites, fully validated within the digital twin before any real-world deployment.
The methodology is structured around the thesis's three axes.
The first axis applies Object-Centric Process Mining (OCPM) to the event logs of the CHU de Toulouse pharmaceutical network, to automatically build and update a digital twin faithful to real processes (stock, orders, lead times, inter-hospital interactions), with a mechanism for continuously detecting deviations between the model and reality.
The second axis mobilises machine learning and deep learning models (XGBoost, random forests, temporal architectures) for demand prediction and anomaly detection from event streams, together with discrete-event simulation to evaluate disruption scenarios (supply shortage, multi-site tension, etc.).
The third axis trains decision policies through deep reinforcement learning (DQN, PPO) in a centralised setting, then extends the approach to multi-agent reinforcement learning (MARL) to coordinate local hospital agents with a central pharmacy agent. All policy training and validation take place within the digital twin before any real-world operational recommendation.

Candidates should hold an engineering degree or a Master's degree (or equivalent) in industrial engineering, computer science, applied mathematics, data science, or artificial intelligence. Strong programming skills (Python) and a solid foundation in machine learning/reinforcement learning are expected, along with a genuine interest in process mining, simulation, and decision support. A good command of scientific English (written and spoken) is required; French is an asset but not mandatory. Candidates should demonstrate autonomy, scientific rigour, and an appetite for interdisciplinary, international work, essential given the Franco-Tunisian co-supervision and the close partnership with a hospital institution.

Application link : https://edd-projets.utoulouse.fr/

• Python
• Anglais
• Machine learning