Thèse Méthodes Guidées par les Données pour l'Ordonnancement Dynamique des Blocs Opératoires H/F Doctorat.Gouv.Fr

Établissement : Université Paris-Saclay GS Sciences de l'ingénierie et des systèmes École doctorale : Sciences Des Systèmes Laboratoire de recherche : Laboratoire Génie Industriel Direction de la thèse : Oualid JOUINI ORCID 000000029498165X Début de la thèse : 2026-10-01 Date limite de candidature : 2026-10-31T23:59:59 Le bloc opératoire est l'une des ressources les plus critiques et les plus coûteuses de l'hôpital, et son fonctionnement est fortement perturbé par l'incertitude sur les durées d'intervention, les durées de séjour post-opératoires et l'arrivée des urgences. L'objectif de cette thèse est de développer des méthodes d'apprentissage automatique pour la recherche opérationnelle, appliquées à l'optimisation dynamique de la planification et de l'ordonnancement des blocs opératoires. Des modèles de machine learning, entraînés sur des données hospitalières historiques et en temps réel, prédiront ces grandeurs stochastiques en quantifiant leur incertitude. Ces prédictions seront ensuite intégrées dans un cadre d'optimisation combinatoire et stochastique pour le (ré)ordonnancement dynamique et multi-objectif des interventions, notamment via des approches « predict-then-optimize » ou d'apprentissage par renforcement. Les méthodes seront validées sur des données hospitalières réelles ou réalistes, avec pour objectif d'améliorer l'utilisation des salles tout en réduisant les heures supplémentaires et les temps d'attente des patients. La thèse sera menée en étroite collaboration avec l'hôpital Pitié-Salpêtrière (AP-HP). Operating rooms are among the most critical and costly resources of hospitals, and their performance directly conditions patients' access to surgical care and the workload of downstream units. Their management is made difficult by the high variability of surgical case durations, the uncertainty of post-operative lengths of stay and the need to accommodate emergency patients alongside elective ones. As a result, operating room schedules are frequently disrupted, leading to overtime, under-utilization of rooms, cancellations and longer patient waiting times. Meanwhile, the growing availability of hospital data opens new opportunities to anticipate these disruptions and to support real-time decision-making. In this context, there is a strong need for data-driven methods combining machine learning and operations research for the dynamic planning and scheduling of operating rooms. The aim of this PhD thesis is to develop:

(1) machine learning models to predict the key stochastic quantities that govern operating room performance, namely surgical case durations, post-operative lengths of stay and emergency arrivals. These models will be trained on historical and real-time hospital data (patient characteristics, type of procedure, surgical team, current state of the operating theatre, etc.) and will provide not only point predictions but also a quantification of their uncertainty, through probabilistic forecasting, quantile regression or conformal prediction methods (Angelopoulos and Bates, 2023).

(2) a data-driven operations research framework coupling these predictions with combinatorial and stochastic optimization for the multi-objective, dynamic (re)scheduling of surgical cases. Both the planning problem (allocation of cases to days and rooms, sequencing within each room) and the online rescheduling problem (decisions triggered by events such as an emergency arrival, a case overrun or a cancellation) will be addressed. Learning-augmented approaches will be investigated, in particular 'predict-then-optimize' schemes, in which predictions are trained with respect to decision quality, and reinforcement learning approaches, in which rescheduling policies are learned through interaction with a simulation model of the operating theatre. The trade-offs between operating room utilization, overtime and patient waiting times will be explicitly accounted for.
Machine learning techniques (e.g., gradient boosting, deep learning, probabilistic forecasting and conformal prediction) will be used to predict surgical case durations, post-operative lengths of stay and emergency arrivals, together with their uncertainty. Operations research tools, mainly combinatorial and stochastic optimization (stochastic programming, Markov decision processes, meta-heuristics), will be used to model and solve the dynamic multi-objective scheduling problems. Learning-augmented approaches ('predict-then-optimize' and reinforcement learning) will be developed to couple prediction and optimization. Discrete-event simulation will be used to evaluate and validate the proposed methods on real or realistic hospital data.