4 Presentations

• 
  01:20 PM - 01:45 PM

  Two-Stage Deep Reinforcement Learning Approach to Solving the Physician Scheduling Problem

  • Imen Masmoudi, presenter, Université du Québec en Abitibi-Témiscamingue
  • Chahid Ahabchane, Université du Québec en Abitibi-Témiscamingue
  • Safa Bhar Layeb, UR-OASIS, National Engineering School of Tunis, University of Tunis El Manar, Tunis, Tunisia, Centre Génie Industriel, Université Toulouse, IMT Mines Albi, Albi, France

  Physician scheduling in emergency departments (EDs) is a complex, NP-hard problem that significantly impacts healthcare efficiency. Traditional manual scheduling is time-consuming, while existing optimization approaches often require extensive recalculation for even minor input changes. This paper explores the application of reinforcement learning (RL) to address the physician scheduling problem (PSP). We propose a novel two-stage training approach, initially utilizing supervised learning to give the agent some basic scheduling logic, followed by deep reinforcement learning (DRL) to optimize for maximal reward while adhering to operational constraints, such as shift coverage, physician availability, and rest requirements defined by the ED of the Rouyn-Noranda, Quebec hospital. We evaluated two agent architectures, one based on dense layers and the other on attention mechanisms, using a real-world dataset. Our results demonstrate that both architectures achieve significant improvements in constraint satisfaction after training. The results suggest that RL offers a promising avenue for generating adaptive and efficient physician schedules. This work contributes a framework for learning optimal scheduling policies from experience, moving beyond static optimization and opening opportunities for incorporating real-time feedback and adaptability within emergency care. Future research will explore alternative RL techniques, improved constraint handling, and the integration of human expertise for a more comprehensive scheduling solution.

• 
  01:45 PM - 02:10 PM

  Master Surgery Scheduling Problem Considering Uncertainty in Time of Stay in Downstream Units

  • Erfaneh Nikzad, presenter, polytechnique montréal
  • Nadia Lahrichi, Polytechnique Montréal

  In this paper, we present a novel approach to the master surgical scheduling problem. Traditionally, this problem involves assigning specialties to specific days and ORs over a planning period, typically spanning 4-6 weeks. Our approach extends this by integrating downstream resources, which significantly impact OR efficiency. We design patient clusters based on their length of stay and surgery duration, incorporating the availability of beds in both the intensive care and surgical units. Additionally, we account for uncertainties in the length of stay. This strategy not only helps prevent surgery cancellations due to bed unavailability but also maximizes the utilization of ORs. To address these uncertainties, we develop a two-stage stochastic programming model. Additionally, we propose a matheuristic algorithm to solve the model efficiently within a reasonable time frame

• 
  02:10 PM - 02:35 PM

  Online Demand Management for Attended Home Delivery

  • Linchuan Wei, presenter, HEC Montreal
  • Yossiri Adulyasak, HEC Montréal
  • Emma Frejinger, DIRO and CIRRELT

  Motivated by a real-world application, we consider the integrated vehicle routing and dynamic booking management problem under uncertain time windows. There is an enduring trade-off between flexibility and operational costs. On the one hand, providing more options increases customer satisfaction and the likelihood of receiving a request. On the other hand, offering many time window options to customers may result in inefficient vehicle routes and schedules, as too many customers may select high-demand time slots. To solve this problem, we incorporate customer choice models to capture customers' preferences for time windows in the vehicle routing problem. The primary decisions include the personalized time window options made available to each customer, the acceptance of booking requests, and vehicle dispatching plans, all made online and dynamically. The problem is modeled using the dynamic programming framework, which is solved by a linear programming approximation. We propose novel policies and demonstrate their effectiveness using real-world data.

• 
  02:35 PM - 03:00 PM

  Process Mining Analysis of COVID-19's Effects on Pulmonary Disease Clinical Pathways

  • Luca Murazzano, presenter, Université Laval
  • Paolo Landa, Université Laval
  • Jean-Baptiste Gartner, Université Laval
  • Mohamed Hakim Raki, Université Laval
  • André Côté, Université Laval

  Pulmonary diseases are major causes of morbidity and mortality globally, significantly impacting healthcare systems due to their chronic nature. In Canada, respiratory diseases such as chronic obstructive pulmonary disease, pneumonia, and lung cancer continue to pose substantial challenges to healthcare management. The COVID‑19 pandemic further exacerbated these challenges, severely disrupting care for pulmonary patients. This study examines the impact of COVID‑19 on the clinical pathways of patients with pulmonary diseases at the Cardiology and Respiratory
  University Hospital in Quebec. Using process mining, we analyze clinical data from the pre‑pandemic (2018‑2019) and pandemic (2020‑2022) periods to assess changes in patient care, focusing on hospitalization and Emergency Room episodes. Key findings indicate a reduction in the number and duration of episodes, with a general streamlining of activities. Notably, patient transitions were expedited during the COVID‑19 era, likely due to the need to manage increased patient loads and minimize infection risks. However, variability in the time between request and completion of
  exams and procedures suggests that the impact on care quality remains unclear. While healthcare systems adapted to the immediate pressures of the pandemic, further research is needed to evaluate the long‑term effects on patient outcomes and care quality. The study concludes with recommendations for optimizing clinical pathways in future public health crises, emphasizing the need for comprehensive data analysis and integrating additional techniques such as machine learning.