INFORMS 2021 Program Book

INFORMS Anaheim 2021

SB15

top of a Multi-layer perceptron model to interpret and examine the underlying clinical features responsible for the classification of Post-Traumatic Headache vs Healthy control patients. The method is able to provide subject-level examination and interpretation. 3 - A Hybrid Computer Simulation Approach to Manage No-Shows in Primary Care Operations Ammar Abdul Motaleb, University of Texas-Arlington, Arlington, TX, United States, Amith Viswanatha, Yuan Zhou, Yan Xiao, Kay Yut Chen, Ayse Gurses, PROMI S. Lab Investigators Patient no-show and late cancellation disrupt the exasperated primary care operations. This practice has adverse ramifications such as decreased clinic resources utilization, increased healthcare costs, among others. To examine the impacts of such disruption on clinic operations and patient satisfactions, this study develops a hybrid computer simulation model that integrates discrete-event simulation (DES) and agent-based simulation (ABS) to represent the flow of patients and micro-level behaviors of clinic personnel. Further, this study designs a set of computer experiments to evaluate the effectiveness of various no-show handling strategies and sheds some lights on its implications in primary care operations management. 4 - Prediction of Inpatient Disaggregate Length of Stay for Heterogeneous Demand Using Machine Learning Algorithms and Survival Analysis Jorge Andrés Acuña, University of South Florida, Tampa, FL, United States, Jose L. Zayas-Castro, Weimar Ardila In the last decades, there has been increased interest in machine learning algorithms and survival analysis to improve hospital performance. Accurate prediction of patient length of stay is a critical metric for healthcare providers and hospital decision-makers. In this talk, we present a framework of prediction models to estimate patients’ disaggregate length of stay. We also study the relationship between the total length of stay and the admission to different care units, such as ICU. Our results provide insights on how to mitigate admission to intensive units and improve patient access to care. SB17 CC Room 202A In Person: Robustness of Neural Networks General Session Chair: Somayeh Sojoudi, University of California-Berkeley, Berkeley, CA, 94530, United States Chair: Brendon Anderson, University of California-Berkeley, Berkeley, CA, 94709-1543, United States 1 - Convex Formulation of Robust Two-layer Neural Network Training Recent work has shown that the training of a two-layer, scalar-output fully- connected neural network with ReLU activations can be reformulated as a finite-dimensional convex program. Leveraging this result, we derive convex optimization approaches to solve the “adversarial training” problem, which aims to train neural networks that are robust to adversarial input perturbations. These convex problems are derived for the cases when hinge loss and squared loss between the network output and the target are used to calculate the training cost. Our work provides an alternative adversarial training method over the current approximation methods, such as Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD). We demonstrate in different experiments that the proposed method achieves higher adversarial robustness than existing training methods. 2 - A Closer Look at Accuracy vs. Robustness Yao-Yuan Yang, University of California, San Diego Current methods for training robust networks lead to a drop in test accuracy, which has led prior works to posit that a robustness-accuracy tradeoff may be inevitable in deep learning. We take a closer look at this phenomenon and first show that real image datasets are actually separated. With this property in mind, we then prove that robustness and accuracy should both be achievable for benchmark datasets through locally Lipschitz functions, and hence, there should be no inherent tradeoff between robustness and accuracy. Through extensive experiments with robustness methods, we argue that the gap between theory and practice arises from two limitations of current methods: either they fail to impose local Lipschitzness or they are insufficiently generalized. Yatong Bai, University of California-Berkeley, Berkeley, CA, United States, Tanmay Gautam, Yu Gai, Somayeh Sojoudi

SB15 CC Room 201C In Person: Advances in Data Analytics for Operations Management and Decision General Session Chair: Yonggab Kim Making, Purdue University, West Lafayette, IL, United States 1 - Drone Delivery Vehicle Routing Problem with Multi-Flight Level Using Gradient Boosting Yonggab Kim, Purdue University, West Lafayette, IL, United States, Hoyoung Jung, Seok Cheon Lee Flight level and delivery efficiency come at a tradeoff. Placing drones higher requires more time, but the higher they are, the less detour they make due to the smaller number of buildings at higher altitudes. We propose a novel vehicle routing problem and solution approach using gradient boosting for multi-flight level drone delivery which aims to minimize delivery completed time. 2 - Interpretable Control with Synthetic Models Yuting Yuan, University of Rochester, NY, Rochester, NY, United States In operational planning problems, organizations collect data, learn the system, and take prompt actions. We identify three potential problems: noise in data, difficulty in counter-factual analysis, and lack of interpretability. To tackle these issues, we propose a new framework that prescribes a data-driven policy regularized by a synthetic model. We demonstrate through experiments that our approach outperforms the benchmark method. 3 - A Dynamic Resilience Management for Deep-Tier Automotive Supply Networks Elham Taghizadeh, Wayne State University, Clinton Township, MI, 48035-5630, United States We propose methods to optimize the resilience of deep tier automotive supply networks. Research confirms that complexity across supplier tiers of automotive supply networks can lead to vastly different network resilience in comparison with simpler supply networks. We integrate network analysis techniques combined with discrete-event simulation informed by secondary data sources and global supply risk databases for improving resilience management. We also demonstrate that optimal resilience strategies across the network. in Healthcare General Session Chair: Nathan B. Gaw, Georgia Institute of Technology, Atlanta, GA, United States 1 - Predicting County-level Pandemic Risk and Relevant Risk Factors Using Machine Learning Kevin Smith, University of Michigan, Ann Arbor, MI, United States, Brian T. Denton, Siqian Shen We aim to determine whether United States (US) counties could be classified for coronavirus disease 2019 (COVID-19)-like disease outcomes using county-level predictive factors and which of those factors are most important to the classification model. We conduct a backward variance inflation factor selection procedure to remove significant multicollinearity among county-level socioeconomic, health, and demographic characteristics. We apply random forests and logistic regression to train models to predict five unique county-level COVID- 19 outcome model scenarios. We compare the results of model scenarios using the Area Under the Receiver Operating Characteristic curve performance measure and report the average of this measure across five stratified cross-validation folds. Our models classify the presence of COVID-19 cases in early outbreak scenarios with excellent discrimination. Socioeconomic factors provide the largest score increases in risk stratification of US counties. 2 - Interpreting Deep Learning Model Predictions Using Shapley Values Jay Shah, Arizona State University, Tempe, AZ, United States ASU-Mayo Center for Innovative Imaging, Tempe, AZ, United States, Catherine Chong, Catherine Chong, Todd Schwedt, SB16 CC Room 201D In Person: Data Science for Complex Data

Todd Schwedt, Visar Berisha, Jing Li, Katherine Ross, Gina Dumkrieger, Jianwei Zhang, Nathan B. Gaw, Simona Nikolova, Teresa Wu, Teresa Wu

More than 2 million people are diagnosed with concussions each year and one of the most common symptoms immediately following a concussive injury is Post- Traumatic Headache. We developed a Shapley value-based approach (SHAP) on

6