Informs Annual Meeting 2017

MD65

INFORMS Houston – 2017

5 - A Risk Hedging Strategy for Wind Power Producers in U.S. Electricity Markets via Financial Risk Exchange Hunyoung Shin, University of Texas at Austin, 10050 Great Hills Trail, Unit 1218, Austin, TX, 78759, United States, hunyoung@utexas.edu Wind power producers (WPPs) participating in forward electricity markets are exposed to real-time (RT) market risks from uncertain generation outputs and RT market prices. To mitigate this joint volume-price risks, we propose a financial instrument referred to as a risk exchange (REX) that enables the WPPs to trade random net payments from uncertain RT prices and MW outputs, after the day- ahead (DA) market is cleared. A negotiation of the REX is analyzed by a bargaining game based on a conflict of interest in determining the REX amounts. Numerical examples show that the REX can reduce RT market risks successfully and encourages the WPPs to sell more energy to the DA market. 370E Predictive Medicine on Electronic Health Record Sponsored: Data Mining Sponsored Session Chair: Kazim Topuz, Oklahoma State University, ktopuz@okstate.edu 1 - Cleaning Challenges and Algorithms for Electronic Health Record Data Zhuqi Miao, Health Data Analyst, Oklahoma State University, 345 Business Building, Stillwater, OK, 74078, United States, zhuqi.miao@okstate.edu, Shrieraam Sathyanarayanan, Elvena Fong, William Paiva Rich health data created via electronic health record (EHR) systems is stimulating the rapid development of EHR-based health research. EHR data can be very “dirty,” but there is little discussion in the literature regarding specific summaries and taxonomies of dirty EHR data and associated cleaning methods. By leveraging Cerner Health Facts, one of the nation’s largest EHR data warehouses, we identified prominent data cleaning challenges and developed algorithms to address these issues. Given the considerable similarities among various EHR systems, it is expected that the findings and algorithms based on Health Facts can be extended to cleaning EHR data in general. 2 - Developing a Synthetic Informative Minority Over-sampling (SIMO) Algorithm Embedded in Support Vector Machines to Learn from Imbalanced Datasets Saeed Piri, Oklahoma State University, School of Industrial Engineering & Management, Engineering North, Stillwater, OK, 74078, United States, saeed.piri@okstate.edu, Dursun Delen, Tieming Liu A dataset is called imbalanced when the distribution of different classes in the data is not similar. While a standard machine learning technique could have a good performance on a balanced dataset, when applied to an imbalanced dataset its performance deteriorates dramatically. In this study, we propose a synthetic informative minority over-sampling (SIMO) algorithm imbedded into support vector machine (SVM). In this algorithm, minority examples close to the SVM decision boundary are over-sampled. We assessed SIMO in comparison to other existing algorithms by using 15 benchmark datasets and it outperformed all of them. We also applied SIMO to a real-word dataset to predict diabetic nephropathy. 3 - Predicting 30-day Hospital Readmissions of COPD: A Data Mining Approach to Study the Effects of COPD Drug Therapies Andrea Blair, Graduate Research Assistant, Oklahoma State University - Center for Health Sciences, 1111 W. 17th Street, Tulsa, OK, 74107, United States, andrea.blair@okstate.edu, Shrieraam Sathyanarayanan Chronic obstructive pulmonary disease (COPD) is a leading cause of 30-day readmissions. Using data extracted from Cerner Health Facts, we built logistic regression, decision tree, and neural network models to predict 30-day readmissions using comorbidities, demographics, and drug therapies. Inpatients with COPD, prescribed one of three common COPD drug therapies (Advair, Symbicort, and Spiriva) or combination of drugs were included in analysis. We used SMOTE (synthetic minority oversampling) to balance the data. Of 52 variables modeled, 28 emerged as important in predicting 30-day readmission for COPD. Decision tree model performed better than other models (AUC DT=0.67, NN=0.66, LR=0.62). MD64

4 - A Decision Support System to Predict Mortality in Sepsis Akash Gupta, Oklahoma State University, 40 South University Place, Stillwater, OK, 747075, United States, akashg@okstate.edu, Tieming Liu, Scott Shepherd, William Paiva We developed a decision tree frame work to predict 28-day mortality among suspected infected patients (58,231 encounters extracted from Cerner’s health fact data) using the vital signs and laboratory tests observed within 24 hours of the admission time. A survival ensemble approach was also used to predict the time to mortality. Our model discriminate (AUROC) non-survival from survival with a accuracy of 0.77 (95% CI = 0.75-0.79). We performed sensitivity analysis to identify the relevant predictors of sepsis mortality. Compare to SOFA, our model Vikrant Vaze, Dartmouth College, 14 Engineering Drive, Murdough Center, Hanover, NH, 03755, United States, vikrant.s.vaze@dartmouth.edu, Kazim Topuz, William Paiva Surgery scheduling should account for uncertainty in the duration of surgical procedures to minimize delays and maximize resource utilization. We develop a data-driven approach to inform the surgery scheduling optimization process under uncertainty 370F Planning Logistics Systems under Uncertainty Sponsored: TSL, Freight Transportation & Logistics Sponsored Session Chair: Michael Hewitt, mhewitt3@luc.edu 1 - Multimodal Capacity Planning with Uncertainty on Contract Fulfillment Mariangela Rosano, PhD Student, Politecnico di Torino, Corso Duca degli Abruzzi 24, Turin, 10129, Italy, mariangela.rosano@polito.it, Guido Perboli, Guido Perboli, Teodor Gabriel Crainic, Teodor Gabriel Crainic Contracting with carriers or third-party logistics firms is a way to secure transportation or warehousing capacity for the next planning period. However, contracts are concluded under uncertainty regarding the demand of loads to be transported, the availability of contracted capacity and of the extra one if needed. The question is how much to contract of each type of capacity to minimize the costs of the contract and of ad-hoc future actions to redress an unfavorable situation. We model this planning problem by extending the Stochastic Variable Cost and Size Bin Packing Problem. Extensive numerical experiments are conducted considering the case of an express courier in a City Logistics environment. 2 - A Learning-based Matheuristic for Stochastic Network Design Teodor Gabriel Crainic, Université du Quebec à Montreal, School of Management, Department of Management & Technology, QC, H3C 3P8, Canada, TeodorGabriel.Crainic@cirrelt.net, Walter Rei, Fatemeh Sarayloo We consider the two-stage stochastic programming formulation of the fixed charge multicommodity capacitated network design problem with uncertain demands. We aim to identify promising design decision variables, common to high-quality solutions. We propose a learning mechanism that systematically extracts such information when scenarios are gradually considered. This mechanism and the information it provides are the core elements of a variable- fixing-based matheuristic. Extensive computational experiment show the efficiently of proposed approach in obtaining high-quality solutions in reasonable computational times, particularly as the instance dimensions increase. 3 - Meso-parametric Value Function Approximation for Dynamic Customer Acceptance Barrett Thomas, University of Iowa, W272 PBB, Iowa City, IA, 52242-1000, United States, barrett-thomas@uiowa.edu, Marlin W. Ulmer We consider a problem in which a dispatcher decides whether to accept delivery requests given resource constraints and the possibility of additional future requests. To solve the problem, we present a novel method of approximate dynamic programming, the offline, meso-parametric value function approximation (VFA). This method combines parametric and non-parametric VFA to estimate future rewards. We compare the approach with the individual VFAs and online rollout algorithms. For a variety of instance settings, we show that the presented method combines the advantages and alleviates the shortcomings of parametric and non-parametric VFAs. uses fewer variables and perform equally well. 5 - Surgery Scheduling under Uncertainty MD65

259