WB2* - Santé / Healthcare 2

Emergency departments (EDs) increasingly face congestion due to rising demand and constrained physician availability. As capacity expansion is often infeasible, effective physician rostering is essential to improve responsiveness and mitigate overcrowding. This work addresses the physician rostering problem by integrating time-varying physician productivity, patient demand and flexible shift design. In particular, the model jointly optimizes shift start times and shift lengths, rather than relying on fixed assignments. Productivity is estimated using clustering methods that identify distinct patterns, which are subsequently adjusted based on assignment type, shift length and individual physician characteristics, thereby capturing dependencies across consecutive shifts. Using real-world data from the emergency department of Policlinico di Milano, we examine clinical outcomes around shift transitions to assess potential fatigue effects. Computational and real-world experiments illustrate improved demand coverage and the generation of realistic schedules.

Nurse rosters are typically optimized deterministically but remain fragile to disruptions. We embed flexibility directly into roster design by modeling shift flips and nurse swaps based on a column generation framework. Results on INRC-II instances show improved access to feasible adjustments and fairness, while preserving cost performance and maintaining computational tractability.

Industrial job cycles involve tasks of varying intensity, duration, and recovery, which complicates the assessment of cumulative muscle fatigue. We propose a discrete fatigue–recovery approach that combines task duration, external load, and maximum voluntary contraction percentage (MVC%) with a Mamdani fuzzy layer to convert intensity, duration, and prior capacity into a Remaining Capacity Index. A second fuzzy module then weights fatigue-recovery effectiveness using literature-derived endurance curves and published endurance–intensity relationships. Simple membership functions map physiological uncertainty to values in [0,1], producing low, moderate and high fatigue categories. Sensitivity analysis identifies an effective 30 R/k window capturing time‑limiting thresholds of sustained effort. Verification against established datasets yields low error (MAE≈0.08, RMSE≈0.10) and strong correlation (r≈0.94) compared to Liu et al 2002. The method balances predictive reliability with physiological interpretability and integrates with Digital Human Modelling to support task planning, shift design and recovery scheduling. Future work will evaluate field deployment and broader validation.

In this paper, we study the stochastic casualty response planning problem and propose a multi-stage stochastic programming model, where initial decisions—such as the location of alternative care facilities (ACFs) and rescue vehicle assignments—are fixed, while patient assignments and allocations of apheresis machines (AM) for blood extraction are updated dynamically as uncertainty unfolds. In this framework, patient demands, blood donor availability, and hospital treatment capacities are treated as uncertain parameters. Our model focuses on deploying newly introduced portable AMs, like the Trima Accel 7, and incorporates both apheresis blood donation and whole blood collection methods, along with platelet transshipments among medical facilities to meet varying demands. This approach aims to improve patient outcomes by optimizing three key performance indicators: timely transport from disaster areas to medical facilities, timely surgeries, and compatible platelet transfusions. By matching platelet age and blood type to injury severity, we reduce the risks associated with mismatched transfusions. To address the scalability challenges in multistage stochastic optimization, we employ stochastic dual dynamic programming (SDDP) and benchmark it against nested Benders decomposition (NBD). To further enhance the efficiency of these algorithms, we develop a wide range of acceleration techniques, including strong cuts, multiple cuts, lower bounding functional valid inequalities (LBFVIs), and warm-up strategies. Effective scenario sampling is achieved through randomized quasi-Monte Carlo (RQMC), while K-means++ clustering is applied for scenario reduction, decreasing the number of scenarios by several orders of magnitude. We carry out extensive computational experiments demonstrating that these enhancements significantly boost the performance of the SDDP algorithm, resulting in a 6,078% reduction in the objective function across 2,500 out-of-sample scenarios and a 40% improvement in the value of the stochastic solution. Finally, a case study from the 2011 Van earthquake in Turkey demonstrates the practical applicability and efficiency of our optimized approach.