Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

SB57

2 - Maximizing Intervention Effectiveness Rongqing Han, University of Southern California, 1111 Wilshire Blvd, Room 406, Los Angeles, CA, 90017, United States, Vishal Gupta, Song-Hee Kim, Hyung Paek Policymakers often seek to roll-out an intervention previously proven effective in a research study, possibly subject to resource constraints. We propose a robust optimization framework to guide this roll-out to maximize intervention effectiveness. Our method uses the types of evidence typically available in a published study to identify a subset of patients for treatment. We prove performance guarantees on the method and present a case study assessing its benefits using data from a large teaching hospital in the context of reducing emergency department visits through case management. 3 - Interpreting Predictive Models for Human-in-the-Loop Analytics Hamsa Sridhar Bastani, Wharton School, Wharton School, Philadelphia, PA, United States, Osbert Bastani, Carolyn Kim Interpretability has become an important issue as machine learning is increasingly used to inform consequential decisions. We propose an approach for interpreting complex “blackbox” models by extracting a decision tree that approximates the model. Our algorithm provably avoids overfitting by actively sampling new training points using the blackbox model. We use this technique to interpret a random forest classifier for predicting diabetes risk. Physicians successfully used our interpretation to discover an unexpected causal issue in the diabetes classifier. 4 - Timing it Right: Balancing Inpatient Congestion versus Readmission Risk at Discharge Pengyi Shi, Purdue University, 403 W. State St, Krannert School of Management, Kran 472, West Lafayette, IN, 47907, United States, Jonathan Helm, Jivan Romain Deglise-Hawkinson, Julian Pan When to discharge a patient plays an important role in hospital patient flow management as well as quality of care and patient outcomes. In this work, we develop and implement a practical decision support tool to aid hospitals in managing the delicate balance between individual readmission risk and ward congestion. Our framework integrates a new prediction model that updates the readmission risk over a patient’s hospital stay with a large-scale Markov Decision Process (MDP) to optimize state-dependent discharge decisions. We overcome both practical challenges in the prediction and the curse of dimensionality in the MDP. Lastly, we discuss the ongoing implementation efforts of this discharge tool. 5 - Flexible FDA Approval Policies Taylor Corcoran, University of California-Los Angeles, 110 Westwood Plaza, Los Angeles, CA, 90024, United States, Elise Long, Fernanda Brava The FDA requires clinical trial evidence that is statistically significant at the 2.5% level when approving novel drugs, but the agency often uses regulatory discretion when interpreting these standards. Factors such as target disease severity, prevalence, and availability of existing therapies are qualitatively considered, yet no quantitative framework is used to evaluate how such characteristics impact approval standards. We propose a novel queueing network model to analyze the drug approval process, which explicitly incorporates these factors, as well as obsolescence among drugs. n SB57 West Bldg 101B Joint Session HAS/Practice Curated: Experiments in Health Care Operations Sponsored: Health Applications Sponsored Session Chair: Hummy Song, Philadelphia, PA, 19104, United States 1 - The Effects of Occupancy Information Hurdles and Physician Admission Decision Noise on Hospital Unit Utilization Song-Hee Kim, University of Southern California, Los Angeles, CA, United States, Jordan D. Tong, Carol Peden Hospital units usually have high demand that exceeds their capacity, requiring physicians to make admission decisions. Under reasonable conditions, the optimal admission policy should depend on the arriving patient’s severity and the occupancy upon a patient’s arrival. In practice, the occupancy is not always readily accessible and there may exist occupancy information hurdles. We recruit physicians and MTurk workers to study how occupancy information hurdles may systematically affect admission decision behavior. We also examine how random error (in the policy selection and policy execution) may drive predictable biases of over- or under-occupied units depending on the system parameters.

n SB55 North Bldg 232C Joint Session Sports/Practice Curated

Sports Analytics I Sponsored: SpORts Sponsored Session Chair: Walter DeGrange, CANA Advisors, Chapel Hill, NC, 27517, United States 1 - Sports Analytics: Making a Difference or Just a Way to get Free Tickets Walter DeGrange, CANA Advisors, 6727 Falconbridge Rd, Chapel Hill, NC, 27517, United States With the recent success of sports teams heavily using analytics (Astros, Eagles, Penguins, Warriors, Leicester City F.C.), does this mean that analytics has gained a foothold in the sports world? And if so, is there a career path that a high school student can use to become a sports analytics professional? This presentation attempts to answer both of these questions and summarizes all the areas of the application of analytics in sports. It also adds another year of professional team results and additional factors. 2 - NHL Expansion Draft Optimization Timothy Chan, University of Toronto, Mechanical and Industrial Engineering, 5 Kings College Road, Toronto, ON, M5S 3G8, Canada, Kyle Booth, Yusuf Shalaby Las Vegas became the 31st team in the National Hockey League in 2017. To build an initial roster, an expansion draft was held where Vegas selected one player from each of the existing teams. Existing teams were allowed to protect a subset of their players from being selected. In this talk, we present integer optimization formulations that model the protection and selection decision problems. We use this optimization framework to investigate how existing teams can reduce Vegas’ team value by making roster moves to improve protection of their assets. We also illustrate how this model can provide Vegas with insight on which teams it may have leverage to extract additional assets through the draft. 3 - How to Play Fantasy Sports Strategically (and Win) Raghav Singal, Columbia University, New York, NY, United States, Martin Haugh Daily Fantasy Sports (DFS) is a multi-billion dollar industry. We provide a coherent framework for constructing DFS portfolios where we explicitly model the behavior of other DFS players. We maximize the expected reward and connect the problem to the mean-variance optimization from finance literature, allowing us to solve our problem using a sequence of binary quadratic programs. We also introduce a Dirichlet-multinomial process for modeling opponents’ behavior, enabling us to predict and embed opponents’ actions in our decision- making in addition to estimating the value of ``insider trading’’ and “collusionö. We demonstrate the value of our framework by participating in DFS contests. n SB56 West Bldg 101A HAS Pierskalla Best Paper Award – I Sponsored: Health Applications Sponsored Session Chair: Jeremy Goldhaber-Fiebert, Stanford, Stanford, CA, United States Co-Chair: Sze-chuan Suen, University of Southern California, Los Angeles, CA, 90089-0193, United States Co-Chair: Margaret L. Brandeau, Stanford University, Stanford, CA, 94305-4026, United States 1 - Bayesian Sequential Learning for Clinical Trials of Multiple Correlated Medical Interventions Ozge Yapar, University of Pennsylvania, Phiadelphia, PA, United States, Stephen Chick, Stephen Chick We propose an approach to clinical trial design that integrates three important, emerging trends: design for cost-effectiveness, which ensures health-economic improvement of new interventions; multi-arm design, which simultaneously compares multiple interventions; and adaptive design, which dynamically adjusts the sample size and allocation of patients to different interventions over the course of the trial. We construct a multi-arm, adaptive sampling policy to maximize the expected population-level benefit and provide numerical results that illustrate the benefits in the context of trials that test multiple combinations of therapies and Phase II/III dose-finding trials.

51