UC Berkeley

Advances in electronic medical records and healthcare databases enable researchers to easily acquire and analyze large amounts of data, and to build data-driven models to improve the system performance. Surgical departments, in particular, utilize a variety of expensive resources, so efficient appointment scheduling and sequencing decisions that minimize patient-surgeon waiting time and the surgeon-operating room idle time substantially reduce costs. We aim to improve the way that schedules are generated by incorporating both dynamically updated data sets, and the opinions of surgeons.

Our research focuses on appointment scheduling of stochastic tasks on a single server where the task durations are challenging to estimate. The task types are known prior to the appointment date but the task duration data is initially limited so that the estimates need to be continuously updated. Appointment scheduling involves both sequencing the tasks and setting the start time of those tasks. Our goal is to develop a data driven appointment scheduling algorithm for sequencing and scheduling tasks. Our research is motivated by a project we have completed with University of California, San Francisco (UCSF) on surgical scheduling where the tasks are the surgical procedures and the server is the operating room.

In Chapter 1 we introduce the appointment scheduling problem with a motivating example of surgical appointment scheduling. We run some simulations to show patient-surgeon waiting time (tardiness) and the surgeon-operating room idle time (earliness) can be reduced by changing the sequence of the procedures and the start times of the procedures. We go over the appointment scheduling literature with various objective functions. We analyze the objective of minimizing expected earliness and tardiness and bound the performance of the commonly used sequencing heuristic based on the standard deviation of procedure duration.

In Chapter 2 we focus on data-driven appointment scheduling. Without making any distributional assumptions we use the empirical distributions of the procedures while computing the objective function which is the expectation of weighted earliness and tardiness. We study the continuity and the convexity of the objective function and the conditions under which there is an integral optimal solution. We briefly go over the methods to optimize the objective function and also constrain the search space containing the minimizer. We develop sequencing heuristics tailored for this problem. Lastly we talk about data selection algorithms if there are categorical features such as name of the surgeon or surgeon estimates about how long the next procedure might take.