Presentation on YouTube.

In dynamic patient scheduling problems, scheduling decisions are taken periodically (daily, weekly… ) and are evaluated on a rolling horizon. As patient arrivals are stochastic, it is important to take into account uncertainty. Machine learning can provide useful tools for extracting knowledge from the historical distribution of patient arrivals to assist scheduling decisions. In this talk, we will present a case study where a machine learning-based algorithm successfully solves a dynamic scheduling problem in the healthcare domain. Scheduling radiation therapy treatment for cancer patients is a complicated task as the treatment requires multi-appointments and subjects to many technical constraints. The problem is even more challenging as the mix of patients with different priorities and deadlines require a smart allocation of treatment resources. We propose a prediction-based approach, where a regression model learns from offline solutions to smartly delay treatments of non-urgent patients to make space for emergency cases. We also demonstrate how our proposed approach supports explainability and interpretability in scheduling decisions using SHAP values.