Sepsis is a life-threatening response to infection that leads to organ failure, tissue damage, and oftentimes, death. Our work leverages historical health data in order to learn treatment strategies for sepsis that result in improved patient outcomes under limited resources. We learn an interpretable, Markov Decision Process (MDP) model of the system, formulate a robust value iteration algorithm, and solve the problem of limited resource allocation for optimal sepsis treatment.

Problem: ED overcrowding disrupts public safety net. Impact of increasing ED entry fee is unknown.

Methodology/Results: In Hong Kong, increasing the ED fee reduced traffic by 6.3%, targeting less-urgent visits. Frequent visitors decreased their visits, reducing patient abandonment.

Managerial implications: Financial access hurdles alleviate healthcare congestion without discouraging urgent visits. Managing external demand complements internal process improvements in ED management.

In many service systems, service times of customers can be correlated with waiting times. Scheduling under such dependency is challenging as a Markovian state description requires keeping track of all customers' waiting history. We propose an approximate dynamic programming algorithm for multi-class scheduling with wait-dependent service times. Our algorithm can generate policies with simple structures and achieve strong performance which we illustrate in a healthcare setting using real data.