Nowadays, healthcare systems are increasingly constrained by limited resources and a rising interest in efficiency. In this context, the French government has implemented new incentive programs to improve logistics activities within hospitals and increase the efficiency of the health system. These leads to cost savings, improved patient care quality, and well-being of nursing staff. Indeed, several collaborative health reforms with different objectives and legal structures have developed in recent years, in particular, with the creation of Territorial Hospital Groups (THG) in 2016. To optimize and rationalize their deployments on the territory, these new restructurings require efficient decision support tools. This thesis was inspired by the opportunity provided by the THG structural project to address the optimization of several problems arising from hospital logistics at the strategic, tactical, and operational decision-making levels. The objective is to determine the optimal scenarios for the allocation, storage, and distribution of products consumed by the care units of the THG. This study investigated two major problems in logistics pooling. On the one hand, to deal with strategic decisions, we are interested in the allocation of products to THG warehouses and stores in order to organize the logistical flows between them and the care units. On the other hand, tactical and operational decisions were jointly modeled from the perspective of a new rich variant of the Inventory Routing Problem (IRP) in a two-echelon, multi-product, and multi-depot system while allowing split deliveries. Unexpected events in the healthcare domain may occur as a result of seasonal fluctuations or epidemics, affecting either of the two investigated problems. Hence, we proposed a fuzzy chance-constrained programming approach to study the relevance of handling uncertainty related to care unit demand and its economic impact. Several methods for the optimization of inventory and product distribution are proposed: an exact method (integer linear programming), a constructive heuristic, and a meta-heuristic approach entitled GVNS (General Variable Neighborhood Search). Different tests were developed on a randomly generated set of instances to demonstrate the performance of the proposed methods.