Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

WB36

3 - Value-based Healthcare Associated Infection Prevention Scheme Using Machine Learning Elioth Sanabria, Columbia University, New York, NY, United States, David D. Yao Healthcare associated infections (HAI) cost 9.8 billion USD per year, making them a major problem for the US healthcare system. We present a machine learning (ML) aided scheme to give patient targeted preventive measures. This scheme aligns the economic incentives of the hospital to give preventive measures to more patients while reducing overall infection costs. We provide probabilistic guarantees on the economic performance of the scheme for any fitted ML model based on a combination of machinery from NP (Neyman Pearson) classification and VC (Vapnik Chervonenkis) theory. We illustrate the methodology with data of hospital admissions between 2009 and 2016 at NewYork Presbyterian Hospital. 4 - On Modeling Patient-flow Dynamics in the Obstetrics and Gynecology Inpatient Unit Jing Dong, Columbia University, New York, NY, 60208, United States, Yue Hu, Ohad Perry We study the patient flow dynamics of the Obstetric and Gynecologic impatient unit of a large teaching hospital in the US. The model we build is easy to calibrate, highly interpretable and captures the main features of patient flow dynamics in the unit. We also study the day-of-the-week effect on the occupancy level and provide key insights on modeling and managing hospital queues. n WB38 North Bldg 225B Games, Learning and Networks Sponsored: Applied Probability Sponsored Session Chair: Lei Ying 1 - Delay Asymptotics and Bounds for Multi-task Parallel Jobs Weina Wang, Carnegie Mellon University, Pittsburgh, PA, United States, Mor Harchol-Balter, Haotian Jiang, Alan Scheller-Wolf, R. Srikant Motivated by file retrieval in cloud storage systems, we study a problem where jobs consist of multiple tasks and a job is completed only when all of its tasks are completed. The goal is to characterize the tail probabilities of job delay. However, this is a difficult problem whereas computing tail probabilities of task delays is relatively easy. By proving that the task delays at the various servers are associated, we show that assuming independence among task delays yields upper bounds on the tail probabilities of job delay. We then further show that the job delay converges to this upper bound in an asymptotic regime where the number of servers grows and the number of tasks in a job is also allowed to grow. 2 - Bayesian Learning with Random Arrivals Randall Berry, Northwestern University, Department of EECS, 2145 Sheridan Road, Evanston, IL, 60208, United States, Tho Le, Vijay Subramanian We consider the impact of uncertainty about agent arrivals on a model of Bayesian observational learning. We study a discrete-time model in which in each time-slot, an agent may randomly arrive. Agents who arrive have the opportunity to buy a given item. If an agent chooses to buy, this action is recorded for subsequent agents. However, the decisions of agents that choose not to buy are not recorded. Hence, if no one buys in a given slot, agents are unaware if this was due to no agent arriving or an agent choosing not to buy. We study the impact of this uncertainty on the emergence of information cascades. 3 - Heavy Traffic Delay Analysis in Data Center Networks Siva Theja Maguluri, Georgia Institute of Technology, Atlanta, GA, 30339, United States, Daniela Hurtado Lange The drift method has been recently developed to study queueing systems in heavy traffic. This method was successfully used to obtain the heavy traffic scaled sum queue lengths in data center networks, even when the so-called complete resource pooling condition is not satisfied, i.e., when there is more than one bottle-neck resource. In this talk, I will present a novel view of the drift method. Using this view, we will present recent developments in obtaining the complete distribution of the queue lengths in heavy-traffic. 4 - Achieving Small Delay in Data Center Networks Carolyn Beck, University of Illinois, Urbana, IL, United States, Dimitris Katselis, R. Srikant, Weina Wang We will consider packet-level and file-level models of data center networks. For both models, we will explore whether simple resource allocation schemes can achieve order-optimal delays.

n WB36 North Bldg 224B Inventory Management II Contributed Session Chair: Emre Berk, Bilkent University, Faculty of Business Administration, Ankara, 06533, Turkeyö 1 - Inventory Model for Fresh Produce Considering Price, Freshness, and Displayed Stock Level Dependent Demand Bhavin J. Shah, Associate Professor, Indian Institute of Management, Indore, Faculty Office # C-206, First Floor, Prabandh Shikhar, Indore - Madhya Pradesh, 453556, India, Hasmukh Gajjar An inventory model is proposed for fresh produce items to determine unit price, cycle time, displayed stock level, and ending inventory to maximize retailer’s total profit. Demand is considered to be deterministic and dependent on price, freshness and displayed stock. Numerical examples are presented to highlight the theoretical results. 2 - Spare Parts Inventory Control with Phase-out Returns and Additive Manufacturing Yayun Jin, North Carolina State University, Raleigh, NC, 27607, United States, Russell Edward King, Donald Paul Warsing, Semra Sebnem Ahiska Manufacturers are concerned with the inventory control of spare parts in the final phase of their service lifecycle. The final phase starts when the specified tooling to produce this type of spare parts is disposed after the final production run, and it ends when the products associated with the spare parts are no longer used. We consider the inventory control with stochastic demand and deterministic returns where the stock is replenished by newly manufactured parts, phase-out returns and repaired defective parts. We formulate this problem as a Markov decision process model and based on the characterization of the optimal policy, we propose newsvendor-like heuristic policies that perform very well. 3 - Newsvendor Model Revisited with Crowdfunding Emre Berk, Bilkent University, Faculty of Business Administration, Ankara, 06533, Turkey I revisit the classical newsvendor model in the presence of financing uncertainty. I consider reward-based P2P lending. The decision maker faces the possibility of not being able to raise enough capital to successfully launch the product. The financing phase is modeled in a two ways - each corresponding to a market segmentation. The impact of financing uncertainty on optimality is examined analytically and numerical examples are provided. n WB37 North Bldg 225A Applied Probability and Healthcare Sponsored: Applied Probability Sponsored Session Chair: Jing Dong, Columbia University, New York, NY, 60208, United States 1 - Time-varying Tandem Queues with Blocking: Comparing Operating Mechanisms in Hospital Networks Noa Zychlinski, Avishai Mandelbaum, Petar Momcilovic We focus on the mechanisms of Blocking After Service (BAS) and Blocking Before Service (BBS), in time-varying many-server queues in tandem. BAS arises in hospitals when patients must remain in their current ward even after completing their treatment, since there is no available bed at the next ward to which they are referred. BBS arises in short procedures such as cataract surgery, cardiac catheterization and hernia repair: the procedure begins only when there is an available bed for the patient in the recovery room. Using reflection analysis, we develop the fluid limit for each mechanism. This gives rise to design/operational insights including a performance comparison between the two mechanisms. 2 - Forecasting Arrivals and Occupancy Levels in an Emergency Department Xiaopei Zhang, Columbia University, New York, NY, 10025, United States, Ward Whitt This is a sequel to Whitt and Zhang (2017), where we developed an aggregate stochastic model of an emergency department. The new work here focuses on forecasting future daily arrival totals and predicting hourly occupancy levels, given recent history. By involving more data, we identify i) long-term trends in both the arrival process and the length-of-stay distributions and ii) dependence among successive daily arrival totals. Both highly structured models and flexible machine learning methods are explored. We find that a SARIMAX time-series model is most effective. We then combine our previous ED model with the arrival prediction to create a real-time predictor for the future ED occupancy levels.

454