Machine learning methods for health care delivery optimization have the potential to improve retention in HIV care, a critical target of global efforts to end the epidemic. However, these methods have not been widely applied to medical record data in low- and middle-income countries. We used an ensemble decision tree approach to predict risk of disengagement from HIV care (missing an appointment by ≥28 days) in Tanzania. Our approach used routine electronic medical records (EMR) from the time of antiretroviral therapy (ART) initiation through 24 months of follow-up for 178 adults (63% female). We compared prediction accuracy when using EMR-based predictors alone and in combination with sociodemographic survey data collected by a research study. Models that included only EMR-based indicators and incorporated changes across past clinical visits achieved a mean accuracy of 75.2% for predicting risk of disengagement in the next 6 months, with a mean sensitivity of 54.7% for targeting the 30% highest-risk individuals. Additionally including survey-based predictors only modestly improved model performance. The most important variables for prediction were time-varying EMR indicators including changes in treatment status, body weight, and WHO clinical stage. Machine learning methods applied to existing EMR data in resource-constrained settings can predict individuals’ future risk of disengagement from HIV care, potentially enabling better targeting and efficiency of interventions to promote retention in care.
A merchant dynamically sets prices in each time period when selling a product over a finite time horizon with a given initial inventory. The merchant utilizes new external information that is observed at the beginning of each time period, whereas the demand function—how the external information and the price jointly impact that single-period demand distribution—is unknown. The merchant’s decision, setting price dynamically, serves dual roles to learn the unknown demand function and to balance inventory with an ultimate objective to maximize the expected cumulative revenue. The main objective of this work is to characterize and provide a full spectrum of relations between the order of optimal expected cumulative revenue achieved in three decision-making regimes: the merchant’s online decision-making regime, a clairvoyant regime with complete knowledge about the demand function, and a deterministic regime in which all the uncertainties are relaxed to the expectations. In the analyses, we derive an unconstrained representation of the optimality gap for generic constrained online learning problems, which renders tractable lower and upper bounds for the expected revenue achieved by dynamic pricing algorithms between different regimes. This analytical framework also inspires the design of two dual-based dynamic pricing algorithms for the clairvoyant and online regimes. This paper was accepted by Hamid Nazerzadeh, data science. Supplemental Material: The online appendix and data are available at https://doi.org/10.1287/mnsc.2023.4963 .
8 March 2023 in Journal of Systems Science and Systems Engineering
Solving Convex Discrete Optimization via Simulation via Stochastic Search Algorithms Many decision-making problems in operations research and management science require the optimization of large-scale complex stochastic systems with high-dimensional discrete decision variables. For a number of applications, the objective function exhibits convexity in the discrete decision variables, or the problem can be transformed into a convex one. In “Gradient-Based Algorithms for Convex Discrete Optimization via Simulation,” Zhang, Zheng, and Lavaei propose provably efficient simulation–optimization algorithms for general convex optimization via simulation problems with high-dimensional discrete decision space. By utilizing the convex structure, the developed algorithms demonstrate a polynomial dependence on the dimension and scale of the decision space. In addition, information-theoretic lower bounds are derived to show the limit of algorithm design. In the case when possibly biased gradient estimators are available, they propose simulation–optimization algorithms to achieve optimality guarantees with a reduced dependence on the dimension.
We build a unified modeling framework for classical online matching problems and emerging online matching problems with three additional practical features: reusable resources, network resources, and decaying rewards. For online matching problems in the unified framework, we provide a unified performance analysis tool for the greedy policy and its simple variants, which we refer to as greedy-like policies. We prove that greedy-like policies can achieve near-optimal performances for online matching problems in the unified framework, where the policy performance is measured by competitive ratios under adversarial environments. We then analyze several representative special classes of online matching problems, which incorporate additional realistic structural assumptions on top of the unified framework. Specifically, we consider online matching problems with each of the following three additional structures: (i) singleton resources with time-decaying rewards; (ii) network resources with accept/reject decisions; and (iii) network resources with interval-type bundles. We show that for these special classes of online matching problems, slight modifications to greedy-like policies can successfully utilize additional structural information to further enhance policy performances. This work may suggest that the greedy policy and its variants, despite its simplicity, can achieve reliable performances for a number of emerging online matching problems. This paper was accepted by J. George Shanthikumar, data science. Funding: The work of D. Simchi-Levi and F. Zhu is partially supported by the Massachusetts Institute of Technology Data Science Laboratory. Supplemental Material: The online appendix is available at https://doi.org/10.1287/mnsc.2023.02588 .
Notably known for its extraordinary thermal and mechanical properties, graphene is a favorable building block in various cutting-edge technologies such as flexible electronics and supercapacitors. However, the almost inevitable existence of defects severely compromises the properties of graphene, and defect prediction is a difficult, yet important, task. Emerging machine learning approaches offer opportunities to predict target properties such as defect distribution by exploiting readily available data, without incurring much experimental cost. Most previous machine learning techniques require the size of training data and predicted material systems of interest to be identical. This limits their broader application, because in practice a newly encountered material system may have a different size compared with the previously observed ones. In this paper, we develop a transferable learning approach for graphene defect prediction, which can be used on graphene with various sizes or shapes not seen in the training data. The proposed approach employs logistic regression and utilizes data on local vibrational energy distributions of small graphene from molecular dynamics simulations, in the hopes that vibrational energy distributions can reflect local structural anomalies. The results show that our machine learning model, trained only with data on smaller graphene, can achieve up to 80% prediction accuracy of defects in larger graphene under different practical metrics. The present research sheds light on scalable graphene defect prediction and opens doors for data-driven defect detection for a broad range of two-dimensional materials.