Integrated Machine Learning and Optimization Frameworks with Applications in Operations Management

Abstract

Incorporation of contextual inference in the optimality analysis of operational problems is a canonical characteristic of data-informed decision making that requires interdisciplinary research. In an attempt to achieve individualization in operations management, we design rigorous and yet practical mechanisms that boost efficiency, restrain uncertainty and elevate real-time decision making through integration of ideas from machine learning and operations research literature.

In our first study, we investigate the decision of whether to admit a patient to a critical care unit which is a crucial operational problem that has significant influence on both hospital performance and patient outcomes. Hospitals currently lack a methodology to selectively admit patients to these units in a way that patient’s individual health metrics can be incorporated while considering the hospital’s operational constraints. We model the problem as a complex loss queueing network with a stochastic model of how long risk-stratified patients spend time in particular units and how they transition between units. A data-driven optimization methodology then approximates an optimal admission control policy for the network of units. While enforcing low levels of patient blocking, we optimize a monotonic dual-threshold admission policy. Our methodology captures utilization and accessibility in a network model of care pathways while supporting the personalized allocation of scarce care resources to the neediest patients. The interesting benefits of admission thresholds that vary by day of week are also examined.

In the second study, we analyze the efficiency of surgical unit operations in the era of big data. The accuracy of surgical case duration predictions is a crucial element in hospital operational performance. We propose a comprehensive methodology that incorporates both structured and unstructured data to generate individualized predictions regarding the overall distribution of surgery durations. Consequently, we investigate methods to incorporate such individualized predictions into operational decision-making. We introduce novel prescriptive models to address optimization under uncertainty in the fundamental surgery appointment scheduling problem by utilizing the multi-dimensional data features available prior to the surgery. Electronic medical records systems provide detailed patient features that enable the prediction of individualized case time distributions; however, existing approaches in this context usually employ only limited, aggregate information, and do not take advantages of these detailed features. We show how the quantile regression forest, can be integrated into three common optimization formulations that capture the stochasticity in addressing this problem, including stochastic optimization, robust optimization and distributionally robust optimization.

In the last part of this dissertation, we provide the first study on online learning problems under stochastic constraints that are "soft", i.e., need to be satisfied with high likelihood. Under a Bayesian framework, we propose and analyze a scheme that provides statistical feasibility guarantees throughout the learning horizon, by using posterior Monte Carlo samples to form sampled constraints that generalize the scenario generation approach commonly used in chance-constrained programming. We demonstrate how our scheme can be integrated into Thompson sampling and illustrate it with an application in online advertisement.