Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

TE58

increase is higher in models that use summary scores compared to those that use all variables as predictors. We conclude that there is important clinical information in the fact that certain fields are missing in EHR data and the development of scoring systems can perform better by considering which variables are missing. 2 - Real-time Prediction Of Sepsis In Hospitalized Adults Using Continuous Bedside Physiological Data Streams Franco van Wyk, University of Tennessee, Knoxville, TN, United States, Anahita Khojandi, Robert L. Davis, Rishikesan Kamaleswaran Sepsis is an acute, life-threatening condition, often acquired in the hospital. Undetected, sepsis can progress to severe sepsis and septic shock, with a risk of death as high as 30% to 80%. Early detection of sepsis can improve patient outcomes. We use a multi-layer machine learning algorithm to analyze continuous, high frequency physiological data, such as vital signs, to identify at risk patients before sepsis onset. In our analysis of a cohort of 1,300 patients, the model only failed to predict 3.16 3.16% of sepsis patients earlier than Systemic Inflammatory Response Syndrome (SIRS) criteria. Sepsis patients were predicted on average 211.47 5.08 minutes earlier than SIRS criteria. 3 - A Machine Learning Model to Predict Adult Inpatient Death, Cardiac Arrest, and ICU Transfer Andres Garcia-Arce, Geisinger, Danville, PA, United States, John Mcilwaine, Jason S. Puckey There is a growing interest in including big data analytics into clinical decision support tools. This project involves the use of electronic records from 79,232 Geisinger patients to create a new early warning score (EWS) using machine learning algorithms. The new model predicts sudden death, cardiac arrest and unplanned transfers to the ICU. This model is expected to outperform other models such as MEWS and NEWS. By having a better EWS, Geisinger clinicians can better plan patient care and hopefully mitigate the effects of adverse events, improving patient outcomes and patient experience. n TE58 West Bldg 101C Joint Session HAS/ICS: Stochastic Modeling and Control in Health Care Sponsored: Health Applications Sponsored Session Chair: Amin Khademi, Clemson University, Central, SC, 29630, United States 1 - Optimal Pooling Schemes for Prevalence Estimation of Emerging Infections Ngoc Nguyen, PhD Candidate, Virginia Tech, 1145 Perry Street, 210 Durham Hall, Blacksburg, VA, 24061, United States, Ebru Korular Bish, Douglas R. Bish Accurate prevalence rate estimation is essential for surveillance of infections and diseases, and is typically conducted via pooled testing. The pooling design, i.e., the number and sizes of testing pools, impacts the estimation accuracy. However, determining an optimal pooling design requires an a priori estimate of the unknown prevalence rate, and can be challenging, especially for emerging infections with limited information. We develop robust optimization models and characterize structural properties of optimal pooling designs for surveillance. Our case study suggests that estimation accuracy can be substantially improved through optimal pooling designs. 2 - Proportional Fairness Model of the Allocation for Heart Transplantation System Farhad Hasankhani, Clemson University, Clemson, SC, United States The allocation process of scarce resources such as donor hearts should be managed to be distributed efficiently and fairly among patients demanding transplant. In order to model fairness in heart allocation, we define utility of each patient as their total life-expectancy and use proportional fairness notion to develop an optimization problem. Objective function is the sum of logarithm of expected utilities of patients in which expected utilities are quadratic functions of the decision variables. The solution of the optimization problem characterizes the structure of the optimal fair allocation rule. 3 - Antiretroviral Therapy Success Huaiyang Zhong, Stanford University, Stanford, CA, 94305, United States Thanks to the success of antiretroviral therapy (ART), HIV has become a chronic disease. Different individual characteristics such as sex, age and race, comorbidities conditions, and drug response history may lead to different optimal choices of ART regimens for a patient in different time periods. Our research aims to provide a personalized HIV treatment plan for patients based on individual characteristics, medical history and other information in patients’ electronic medical records.

n TE56 West Bldg 101A Decision-making in Drug Research and Development Sponsored: Health Applications Sponsored Session Chair: Alba Rojas-Cordova, Southern Methodist University, Dallas, TX, 75219, United States 1 - An Approximation Approach for Response Adaptive Clinical Trial Design Vishal Ahuja, Southern Methodist University, Cox School of Business, P.O. Box 750333, Dallas, TX, 75275, United States, John R. Birge Multi-armed bandit problems exemplify the exploration vs. exploitation tradeoff. For many practical problem, the state space is intractably large, rendering exact solution approaches impractical. We propose a novel approximation approach that combines grid-based techniques, simulation, and methods to improve approximation accuracy, to obtain near-optimal solutions with minimal added computational burden. 2 - Drug Shortages and Pharmaceutical Supply Chain Reliability Emily L. Tucker, University of Michigan, Ann Arbor, MI, 48105, United States, Mark S. Daskin, Burgunda V. Sweet, Wallace J. Hopp A major barrier to effective patient care is drug availability. Shortages within the United States are prevalent and are largely driven by supply chain resiliency decisions. We present a new model of pharmaceutical supply chain reliability to analyze the effect of design decisions on drug availability. We discuss competitive extensions. 3 - Analysis of Medicines Shortages in Europe Vincent Hargaden, University College Dublin, School of Mechanical & Materials Engineering, Engineering & Materials Science Centre, Belfield, DUBLIN 4, Ireland, Rachel Ward, Eoghan O’Reilly The issue of medicines shortages across Europe has been receiving increased attention recently from governments, regulatory agencies and healthcare providers. We report on the work of a European Union research project “European Medicines Shortages Research Network - Addressing Supply Problems to Patients”, funded by the Cooperation in Science and Technology (COST) program. This talk will focus on approaches (e.g. system dynamics) and analysis of the downstream supply chain issues which lead to shortages at both hospital and community pharmacies. 4 - Adaptive Trial Supply Optimization Wei-An Chen, Purdue University, Weldon School of Biomedical Engineering, MJIS Building, Room 1052, West Lafayette, IN, 47907-2032, United States, Nan Kong As adaptive clinical trials receive growing attention, they incur additional challenges to drug supply chain management. Targeting on commonly-used adaptation schemes, size re-estimation and dosage dropping, we developed a stochastic production-inventory-distribution optimization model under participant enrollment and trial size uncertainty, which addresses important aspects such as trial span, resupply policy, drug wastage. Our numerical study shows our model can deliver promising operational responses from the supply chain aspect to the trial design adaptiveness. n TE57 West Bldg 101B Joint Session HAS/DM: Modeling for Complex Conditions: Acute and Chronic Sponsored: Health Applications Sponsored Session Chair: Julie Simmons Ivy, North Carolina State University, Raleigh, NC, 27695-7906, United States 1 - Handling Missing Data in Severity of Illness Score Development: The Knowledge of Missing Information Combined with Imputation Improves Predictive Power Joseph Kapena Agor, North Carolina State University, Raleigh, NC,

27606, United States, Osman Ozaltin, Julie Simmons Ivy, Muge Capan, Ryan Arnold, Santiago Romero-Brufau

The objective of this work is to quantify the impact of missing and imputed variables on the performance of prediction models used in the development of sepsis-related severity of illness scoring systems using EHR data. We find that there is a significant increase in performance when moving from models that do not indicate missing information to those that did (p-value<0.001). This

399