Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

SA71

To achieve better health outcomes, early detection of ineffective treatment plays a key role in developing optimal treatment plans. We formulated a personalized treatment monitoring and switching problem for chronic depression as a Markov decision process, and estimated the individual treatment effects on disease transitions. Considering the tradeoff between exploration and exploitation in solving the MDP, optimal treatment policies were obtained and analyzed using simulated data. This work is a starting point to enable optimal depression treatment selection using a decision support system. 2 - Detect Depression from Communication: How Computer Vision, Signal Processing, and Sentiment Analysis Join Forces Aven Samareh, University of Washington, 4324 8th ave NE, D7, Seattle, WA, 98105, United States, Yan Jin, Zhangyang Wang, Xiangyu Chang, Shuai Huang Depression will leave recognizable markers in patient’s vocal acoustic, linguistic, and facial patterns, all of which have demonstrated increasing promise on evaluating and predicting patient’s mental condition in a more objective way. We developed a multi-modality prediction model to combine the audio, video, and text modalities, to identify the biomarkers that are predictive of depression with consideration of gender differences. We identified promising biomarkers from successive search on feature extraction analysis for each modality. 3 - Forecasting the Demand for Mobile Clinic Service Based on Demographic and Clinic Data Bilal Majeed, University of Houston, Houston, TX, 77054, United States, Jiming Peng, Ying Lin Demand forecasting plays an important role in the deployment of mobile clinic services as it can help a mobile clinic to maximize its coverage under limited resource. In this talk, we present a new forecasting model to predict the delinquency rate in census tracts based on the clinic and the demographic data. For this, we first develop some associations between the delinquency data in census tracts and school zones. Then we combine semi-supervised learning and convex optimization to build up a forecasting model for the delinquency rates in all the census tracts. A case study in Harris County will be reported to demonstrate the efficacy of the new model and technique. n SA71 West Bldg 106C Joint Session ICS/IOS-Uncertain: Stochastic and Distributionally Robust Optimization Sponsored: Computing Sponsored Session Chair: Hamed Rahimian, Northwestern University, 2145 Sheridan Rd, Evanson, IL, 6208, United States 1 - Distributionally Robust TSP with Wasserstein Distance Mehdi Behroozi, Northeastern University, Department of Mech. & Ind. Engineering, 334 Snell Engineering Center, Boston, MA, 02115, United States, John Gunnar Carlsson, Kresimir Mihic Motivated by a districting problem in multi-vehicle routing, we consider a distributionally robust version of the travelling salesman problem in which we compute the worst-case spatial distribution of demand against all distributions whose Wasserstein distance to an observed demand distribution is bounded from above. This allows us to circumvent common overestimation that arises when other procedures are used, such as fixing the center of mass and the covariance matrix of the distribution. Numerical experiments confirm that our new approach is useful when used in a decision support tool for dividing a territory into service districts for a fleet of vehicles when limited data is available. 2 - Distributionally Robust Optimization with Chance Constraints Using Wasserstein Metric Ran Ji, George Mason University, 4400 University Dr. MS 4A6, Fairfax, VA, 22030, United States, Miguel Lejeune We study distributionally robust chance-constrained optimization problems DRCCP with Wassertein metric under two types of uncertainties (uncertain probabilities and continuum of realizations). For the case of uncertain probabilities (resp. continuum of realizations), we propose a set of deterministic mixed-integer linear programming (resp. second-order cone programming) inequalities to reformulate DRCCP. We transform the chance constraints to the expectation ones via indicator function, then leverage the convex duality to reformulate the expectation constraints under Wasserstein ambiguity set. We derive valid inequalities to enhance the computational efficiency. 3 - Two-stage Distributionally Robust Mixed Integer Program Manish Bansal, Virginia Tech., 227 Durham Hall, 1145 Perry Street, Blacksburg, VA, 24060, United States In this talk, we present our recent advances for solving two-stage distributionally robust mixed integer program and its variants.

n SA69 West Bldg 106A Joint Session QSR/Practice Curated: Data Analytics in Healthcare Applications Sponsored: Quality, Statistics and Reliability Sponsored Session Chair: Dongping Du, Texas Tech University, Lubbock, TX, 79409, United States 1 - Network-based Analysis of Multivariate Sensory Data for Real-time Monitoring of Cardiac Patients Chen Kan, University of Texas at Arlington, Arlington, TX, United States In this paper, a new network model is developed to integrate multivariate sensory data forreal-time detection of abnormal cardiac patterns. Each cycle of the high- dimensionalsignal is mapped as a weighted network. Structural variations among networks arecharacterized by optimal network alignment, which are then incorporated into anexponentially weighted moving average control chart for in- situ monitoring. Experimentalresults show that the proposed network model effectively and efficiently capturesincipient changes of cardiac activity for the early identification of disease patterns 2 - Atrial Fibrillation Source Identification Using Multinomial Distribution and Maximum Likelihood Estimation Dongping Du, Texas Tech University, 2500 Broadway, P.O. Box 43061, Lubbock, TX, 79409, United States, Amirhossein Koneshloo Atrial Fibrillation (AF) is the most common sustained cardiac rhythm disorder. Identifying AF triggers is essential for efficient treatment design. However, due to the uncertainty and randomness associated with electrical wave propagation, determining the trigger locations is challenging. In this study, we develop a robust probabilistic method to track the electrical wave conduction in AF in the presence of uncertainty. The method provides confidence regions that successfully capture AF sources. The confidence bounds can be valuable references for AF ablation design and planning. 3 - A Mixed Effects Multi-task Learning Model for Parkinson’s Disease Monitoring Using Smartphones Hyunsoo Yoon, Arizona State University, 699 S. Mill Ave, Tempe, AZ, United States, Jing Li Built-in sensors in smartphones can conveniently collect the activities of PD patients. The cost-effectiveness of this method allows patient conditions to be assessed frequently for more effective treatment. However, due to patient heterogeneity, patient-specific models are needed but face the challenge of limited data regarding the Unified PD rating Scale representing disease severity. We propose a multi-task learning approach, MEMTL, which can utilize both general population and patient-specific patterns to improve prediction performance. The unique features of MEMTL include detection of early signs of PD and accurate prediction of PD progression. 4 - Subspace Learning and Representation of Dynamic Coordination Pattern Across Multiple Joints During Movement for Chronic Ankle Instability Shaodi Qian, Northeastern University, Boston, MA, 02115, United States, Sheng-Che Yen, Eric Folmar, Chun-An Chou Ankle sprains and instability are major public health concerns. Up to 70% of individuals don’t fully recover from a single ankle sprain and develop chronic ankle instability (CAI). The diagnosis of CAI is usually based on self-report rather than objective biomechanical measures. In this study, we propose a new analytics method to identify discriminative patterns in the coordination among bilateral hip, knee, and ankle joints between individuals with CAI and health cohorts. A subspace learning algorithm is developed to estimate the multi-joint coordination patterns. The computational results show >80% classification accuracy with the identified patterns using a SVM classifier.

n SA70 West Bldg 106B

Joint session QSR/DM: Challenges and Novel Solutions in Multi-Source Data Integration Sponsored: Quality, Statistics and Reliability Sponsored Session Chair: Ying Lin 1 - Personalized Treatment Monitoring and Switching Policies for Chronic Depression Mutita Siriruchatanon, University of Washington, Seattle, WA, United States, Shan Liu

27