Two Essays on Data-driven Decision Making: Healthcare Analytics and Contextual Online Learning

數據驅動決策: 醫療分析與情景在綫學習

Student thesis: Doctoral Thesis

View graph of relations


Related Research Unit(s)


Awarding Institution
Award date27 Aug 2021


In this thesis, we study two data-driven decision-making problems, and present the results in two essays. In the first essay, we study the problem of patient prioritization in the emergency department (ED). To understand how ED decision makers choose the next patient for treatment, we estimate a discrete choice model and find that ED decision makers apply urgency-specific delay-dependent prioritization. Moreover, when ED blocking level is sufficiently low, admit patients-who need further care at inpatient units-are prioritized over discharge patients for high acuity patients, whereas disposition does not affect the prioritization of middle-to-low acuity patients. When the ED blocking level becomes sufficiently high, decision makers start to prioritize patients who are less likely to be admitted after treatment at ED, in an effort to avoid further blocking the ED. We then analyze a stylized model to explain the rationale behind decision makers' prioritization behavior when the ED faces an increasing risk of being blocked. We also investigate the impact of such prioritization behavior on ED operational performances and show how to leverage our findings to improve ED waiting time prediction. By testing and highlighting the central role of decision makers' patient prioritization behaviors, our work advances the understanding of ED operations and patient flow.

In the second essay, we consider a contextual online learning (multi-armed bandit) problem with high-dimensional covariates and decisions. The reward function to learn does not have a particular parametric form. The optimal regret shown in the literature suffers from the curse of dimensionality. In reducing the dimensionality effect of the covariate, we take advantage of the sparsity structure of the covariate and propose a variable selection algorithm called BV-LASSO, which incorporates novel ideas such as binning and voting to apply LASSO (Least Absolute Shrinkage and Selection Operator) to nonparametric settings. The regret achieved by our algorithm matches the optimal regret when the sparsity structure is known in advance. And thus, the regret cannot be improved. In reducing the dimensionality effect of the decision, we impose the curvature structure and propose an algorithm based on stochastic approximation and binning. The regret achieved does not depend on the dimension of the decision. Our algorithms may serve as a general recipe to achieve dimension reduction in nonparametric settings.

    Research areas

  • Patient Prioritization, ED Blocking, Discrete-Choice Model, MDP, Simulation, Waiting Time Prediction, Contextual Bandits, Nonparametric Variable Selection, LASSO