INFORMS 2021 Program Book

INFORMS Anaheim 2021

SC15

SC15 CC Room 201C In Person: Advances in Medical Informatics General Session Chair: Xiaochen Xian, University of Florida, Gainesville, FL, 32608- 1012, United States 1 - Functional Regression Based Quantification of Interruptive Effects for Resuscitation Events in Hospital Emergency Departments Xiaochen Xian, University of Florida, Gainesville, FL, 53719-2450, United States Large-scale service systems experience disruptive events that cause diversions of resources, which poses a challenge in maintaining high-quality services. The paper focuses on modeling data-driven large-scale service systems and proposing a metric to quantify the disruption effects of certain interruptive events. A functional regression based modeling and analysis scheme is applied to accurately model the system service status and fully understand possible disruptive effects. The effectiveness of the proposed method is validated via a real case study from ED of a hospital. 2 - Optimal Intervention Portfolio to Improve Health Equity Response in the Coronavirus Pandemic Yueran Zhuo The Coronavirus (COVID-19) pandemic hit the United States tremendously with the shock heavily fallen into the underrepresented communities. The purpose of this research is to help these organizations find the optimal strategy to serve the underrepresented communities through the COVID pandemic. Specifically, we aim to help them understand the difference between underrepresented communities and regular communities regarding their need for preventive awareness, testing and medical/supportive care. This will make a step to our ultimate goal of generalizing these findings to any health equity response program. 3 - Synthesizing Data-driven Sepsis Treatment Strategy Using Longitudinal EHR Data Akash Gupta, California State University, Northridge, CA, 91324, United States, Michael Lash, Senthil Nachimuthu Sepsis is one of the leading causes of death in Intensive Care Units (ICU). The strategy for treating sepsis involves the infusion of intravenous (IV) fluids and administration of antibiotics. Because of the rapid change in patient’s health, determining the optimal quantity of IV fluids is a challenging problem. In this study, we capture the longitudinal EHR data to develop data-driven treatment strategies. 4 - Atomic Clique: A Novel Network Model to Analyze Comorbidity Progression Parisa Sahraeian, Oklahoma State University, Stillwater, OK, United States Detection and characterization of comorbidity progression is an invaluable decision aid and a prominent challenge in healthcare research and practice. Comorbidity progression can be modeled as temporal disease networks (TDNs). The objective of this study is to detect comorbidity progression patterns among TDNs. In this regard, a new network model, Atomic Clique, and an associated optimization problem, Atomic Clique Partition (ACP) problem, were proposed. The effectiveness of the model was demonstrated using two case studies on C. Diff and stroke. 5 - Effect of Clinical Measurement Errors on Tuberculosis Treatment Outcomes Prediction Maryam Kheirandish, University of Arkansas, Fayetteville, AR, United States, Shengfan Zhang In this research, the effect of predictors’ measurement error on Tuberculosis (TB) treatment outcomes prediction is investigated. Since there are no biomarkers to assess progression of TB burden in patients, many studies developed prediction models to predict TB treatment outcomes using clinical data to help to assess the effectiveness of treatment plans. Although RF models are shown in many studies that perform well in treatment outcomes prediction, there are significant measurement errors and heterogeneity in laboratory test results which affect reliability of these models. This study explores how these errors affect performance of RF models and how these effects could be neutralized to achieve more reliable predictions.

SC16 CC Room 201D In Person: Optimization in Machine Learning General Session Chair: Moontae Lee, University of Illinois at Chicago, Chicago, IL, 60607, United States 1 - Causal Inference for Panel Data with General Treatment Patterns Tianyi Peng, MIT, Cambridge, MA, United States, Andrew A. Li, Vivek Farias We consider the problem of causal inference for panel data with general treatment patterns, a paradigm with broad applications in areas ranging from program evaluation to e-commerce. We propose a novel treatment effect estimator for this problem that we show to be rate-optimal and asymptotically normal under general conditions on the treatment pattern. Our work thus generalizes the synthetic control paradigm to allow for general treatment patterns. Our recovery guarantees are the first of their type in this general setting. Computational experiments with our estimator on synthetic and real-world data show a substantial advantage over competing matrix completion based estimators. 2 - Generalized and Scalable Optimal Sparse Decision Trees Chudi Zhong, Duke University, NC, United States With the widespread use of machine learning, the importance of interpretability has become clear for high-stake decisions. In this talk, I will focus on a fundamental and important problem in the field of interpretable machine learning: optimal sparse decision trees. We propose an algorithm that produces optimal sparse binary-split classification trees through a special combination of branch-and-bound and dynamic programming. It leverages several important theorems to reduce the size of the search space. It generalizes decision tree optimization to handle various objectives including F-score and AUC convex hull and exposes a high degree of computational reuse when modeling continuous features. 3 - On-the-fly Rectification for Robust Large-vocabulary Topic Inference Moontae Lee, University of Illinois Chicago, Chicago, IL, United States Co-occurrence statistics are powerfully informative. By transforming unsupervised learning into decompositions of co-occurrence, spectral algorithms provide transparent and efficient algorithms for posterior inference such as latent topic analysis and community detection. As object vocabularies grow, however, it becomes rapidly more expensive to store and run inference algorithms on co- occurrence statistics. Rectifying co-occurrence, the key process to uphold model assumptions, becomes increasingly more vital in the presence of rare terms, but current techniques cannot scale to large vocabularies. We demystify previously unknown theories behind the rectification, and then we propose novel approaches that simultaneously compress and rectify co-occurrence statistics, scaling gracefully with the size of vocabulary and the dimension of latent space. 4 - Surrogate “Level-Based” Lagrangian Relaxation for MILP Problems Mikhail A. Bragin, Assistant Research Professor, University of Connecticut, Storrs, CT, United States Combinatorial optimization plays a prominent role in Operations Research. To efficiently solve “separable” combinatorial problems, we developed a decomposition and coordination “level-based” surrogate Lagrangian relaxation method with adaptive adjustment of the “level” estimate of the optimal dual value for faster convergence; the key is faster multiplier “oscillation detection” based on a novel auxiliary “dual-convergence-feasibility” problem. Testing results for generalized assignment problems (GAP) demonstrate high computational efficiency and high solution quality (e.g., with the cost of 97825 within 20 min. for the GAP instance d201600) as well as stability and robustness.

20