Department of Industrial Engineering & Decision Analytics - Learning to Save Lives: Optimal Allocation of Drones for Out-of-Hospital Cardiac Arrests

We study the problem of allocating drones to deliver automated external defibrillators (AEDs) for out-of-hospital cardiac arrests (OHCAs). A key challenge is that the time between the onset of cardiac arrest and the emergency call is unobservable and varies across locations, yet it plays a critical role in determining patient survival. This uncertainty implies that the survival rate curve at each demand location is not known a priori and must be learned through deployment. At the same time, policy decisions must maximize the number of lives saved in real time. This creates a fundamental exploration–exploitation tradeoff, where the planner must learn local survival dynamics while simultaneously optimizing rescue outcomes. We show that the problem can be formulated as a combinatorial multi-armed bandit (CMAB) with latent structural dependencies.

We show that standard CMAB algorithms, including UCB and Thompson sampling, perform poorly under realistic finite-horizon conditions. To understand the source of this behavior, we develop new finite-time regret bounds with explicit constants for a range of common explorative and exploitative policies in a simplified setting. These bounds reveal the structure of the exploration–exploitation tradeoff and clarify how it evolves over time. Building on these insights, we design a simple modified greedy policy that achieves an effective balance between learning and immediate performance in practical, finite-horizon environments. In computational experiments using large-scale, data-driven OHCA instances, this policy can save up to 20 percent more lives than the best-performing standard bandit algorithms, while remaining computationally efficient and scalable.