Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

TA58

2 - Cost-effectiveness and Decision Analysis of Genetic Testing in Cholesterol Treatment Planning Wesley J. Marrero, University of Michigan, Ann Arbor, MI, United States, Mariel Sofia Lavieri, Rodney A. Hayward, Suzanne C. Butler, Amit Khera, Sekar Kathiresan, James Burke, Jeremy B. Sussman We present a simulation-based framework to estimate the risk of heart attacks due to clinical and genetic factors. Additionally, we develop cholesterol treatment plans using risk thresholds (current practice) and a Markov decision process (MDP). By simulating the health status of patients, we determined the cost- effectiveness of genetic testing to guide cholesterol treatment. 3 - Statin Initiation Decision Modeling for Prediabetes Patients Shengfan Zhang, University of Arkansas, 4207 Bell Engineering Center, Department of Industrial Engineering, Fayetteville, AR, 72701, United States, Muhenned Abdulsahib While there is much research on statin initiation policies on prevention of heart disease for diabetes, studies on statin initiation policies for prediabetic patients are limited. The goal of this research is to examine the tradeoffs between the risk of heart disease and risk of diabetes on the decision of statin therapy for prediabetes patients. We develop an optimal statin initiation policy to meet both the need to control cholesterol levels and the need to minimize the risk of diabetes, which will provide insights for future treatment guidelines for prediabetes. 4 - Modeling Comprehensive Medication Reviews for Complex Patients in Community Pharmacies Kathryn N. Smith, North Carolina State University, Raleigh, NC, 27606, United States, Julie Simmons Ivy, Anita Vila-Parrish Adherence to long-term medication therapies is approximately 50% in developed countries. Patient and provider engagement has been found to be a contributor to improved adherence. One enhanced service aimed at engaging patients is a comprehensive medication review (CMR) provided by pharmacists. CMRs allow a pharmacist to identify drug therapy problems that may be interfering with adherence. In order to incorporate CMRs into the workflow, pharmacies must streamline the CMR process and prioritize complex patients. We developed a simulation model and a dynamic programming model to analyze the integration of the operational and clinical workflows and determine how to prioritize patients. n TA58 West Bldg 101C Joint Session HAS/Practice Curated: Healthcare Analytics II Sponsored: Health Applications Sponsored Session Chair: Pinar Keskinocak, Georgia Institute of Technology, Atlanta, GA, 30332, United States Co-Chair: Seyma Guven-Kocak, Georgia Institute of Technology, Atlanta, GA, 30340, United States 1 - One-class Adaptive Resonance for Radiotherapy Quality Assurance Dionne Aleman, University of Toronto, 5 King’s College Road, Toronto, ON, M5S 3G8, Canada, Hootan Kamran Habibkhani, Chris McIntosh, Thomas Purdie Radiotherapy (RT) treatment plans, once designed, must undergo iterative refinements until they pass quality assurance (QA). Machine learning has promised to help automate such QA processes by learning from past data. However, RT plans are usually not recorded until they passed QA and are sent for delivery. Therefore, a clinically-recorded RT dataset is usually imbalanced in favour of the easily-recordable high-quality plans, and binary classifiers struggle to learn from it. Therefore, we develop an adaptive neural network for one-class classification. Our algorithm outperforms one-class SVMs and autoencoders on two RT datasets for breast and prostate patients. 2 - Updating Risk Assessment Tools for Unplanned Extubation in Pediatric Critical Care Patients Zihao Li, Georgia Institute of Technology, Atlanta, GA, United States, Pinar Keskinocak, Atul Vats The Risk assessment score (RAS) and revised RAS (rRAS) are scoring tools based on consensus opinion to assess pediatric patients’ risk of unplanned extubation (UPE). We reviewed records of 5,749 patients from 2013 to 2017 in 5 ICUs and 2 freestanding children’s hospitals and used logistic regression to improve the score. The new model identifies UPE history, weight, age, oral secretion and ventilation length as significant risk factors. The incident rates for low, moderate, high, and extreme risk patients are 0.69, 0.46, 1.14 and 3.01 per 100 vent days, respectively. The new model has a stronger association with UPE occurrence for patients with extreme risk compared to the RAS and rRAS.

3 - Calling for Care? The Risky Proposition of Teletriage for Healthcare Demand Management Ozden Engin Cakici, American University, Washington, DC, United States, Alex Mills We investigate the effect of adding teletriage to a healthcare system with traditional or open access primary care and an Emergency Department (ED). Using a partially observable Markov decision process model, we find that while teletriage would benefit patients, it could be costly for the payer and even increase ED usage. We conclude by providing conditions under which teletriage would be beneficial. 4 - Dynamic Personalized Patient Classification via Learning Progression in Chronic Diseases: Application to Glaucoma Esmaeil Keyvanshokooh, University of Michigan, Ann Arbor, Ann Arbor, MI, 48108-1020, United States, Mark P. Van Oyen, Joshua Stein, Mariel Sofia Lavieri, Chris Andrews We design a dynamic and personalized classification method for classifying a patient with Glaucoma at each visit as either a “fast” or “controlled” progressor. “Fast refers to relatively rapid deterioration with respect to visual field mean deviation. To this aim, we combine a random forest algorithm with a classification method. We also develop online learning methods to help manage a patient’s progression. Joint Session HAS/DM: Aged Care Analytics: Models, Methods and Applications: Part I Sponsored: Health Applications Sponsored Session Chair: Mingyang Li, Tampa, FL, 33647, United States Co-Chair: Nan Kong, Purdue University, West Lafayette, IN, 47906- 2032, United States 1 - Skilled Nursing Facility Service Utilization Modeling and Staffing Evaluation with Competing Discharge Disposition Nazmus Sakib, University of South Florida, Tampa, FL, United States, Xuxue Sun, Nan Kong, Chris Masterson, Hongdao Meng, Kathryn Hyer, Mingyang Li Skilled nursing facilities (SNFs) are major long-term care settings for older adults. It is important to accurately model SNF service demands so that an appropriate staffing level can be determined to ensure high quality care at a reduced cost. This work proposes a data analytics integrated simulation framework to evaluate the service demands of SNF residents and further investigate cost-effective staffing decisions under various resident census and acuity scenarios. The proposed work will serve as a decision support toolset for SNF administrators to improve their staffing decisions. 2 - A Novel Privacy-preserving Positive Transfer Learning Approach for Telemonitoring of Parkinson’s Disease Hyunsoo Yoon, Arizona State University, 699 S. Mill Ave, Tempe, AZ, 85281, United States, Jing Li Telemonitoring of voice signals of Parkinson’s Disease (PD) patients using an At- Home Testing Device is a cost-effective logistically-convenient way to monitor disease progression and optimize treatment. Key challenges in telemonitoring are patient heterogeneity and limited data for each patient. We propose a transfer learning approach, P3TL, which can leverage other patients’ information when building a predictive model for a target patient. The unique features of P3TL include intelligent selection of patients to transfer to avoid negative transfer and maintain patient privacy preservation. 3 - A Novel Transfer Learning Model for Alzheimer’s Disease(ad) Early Detection Using Incomplete Multimodality Image Data Xiaonan Liu, Arizona State University, Tempe, AZ, United States, Kewei Chen, Teresa Wu, David Weidman, Fleming Lure, Jing Li Early detection of AD is critically important for treatment of this devastating disease. Use of multimodality images has been shown to hold great promise. The challenge is that images of different modalities are not universally available to all patients due to cost and accessibility constraints. We propose a novel incomplete- modality transfer learning (IMTL) model that learns a diagnostic model for each patient sub-cohort with the same availability of image modalities while coupling the model training processes to allow knowledge transfer. IMTL achieves significantly better accuracy than existing single learning models on datasets collected by the AD Neuroimaging Initiative (ADNI). n TA59 West Bldg 102A

264