Search

Weaning patients from mechanical ventilators is a critical decision in intensive care units (ICUs), especially during the COVID-19 pandemic. In this study, we aim to improve the extubation decision for ventilated patients by exploiting data analytics and predictive information. We develop a discrete-time, finite-horizon Markov decision process (MDP) with predictions on future information to support the extubation decision. We characterize the structure of the optimal policy and provide important insights into how predictive information can lead to different decision protocols. Using a comprehensive data set from an ICU in a tertiary hospital in Singapore, we compare the performance of different policies and demonstrate that incorporating predictive information can reduce ICU length of stay (LOS) by up to 9.4% and, simultaneously, decrease the failure rate of ventilated patients by up to 18.9%. The benefits are more significant for patients with poor initial conditions. Furthermore, simply optimizing LOS using a classical MDP model without incorporating predictive information leads to an increased failure rate of ventilated patients by up to 6%. Using publicly available data on critical COVID-19 patients, we show that applying the extubation protocols using predictive information can improve ICU throughput by up to 9.0%.

The rapid growth of the on-demand economy offers unprecedented convenience for customers and companies, by transforming the traditional approach to labor management from permanent employees to on-demand employees. At the same time, this shift presents new challenges for the management of an on-demand labor force.In this paper, we study this new phenomenon and its impact on labor market efficiency. We consider employers with uncertain and possibly time-varying demands and a platform operator who provides on-demand staffing services. We first characterize system-wise optimal staffing strategies from the perspective of a central planner, then propose a novel fill rate-based allocation policy that enables the on-demand workforce to be shared efficiency among individual employers. Next, we propose a type of incentive contracts based on fill-rate guarantees, and show that our contracts can induce system-wise optimality in decentralized systems. Our result provides a win-win-win solution to the sharing economy: Individual employers and the platform operator share the benefits of on-demand staffing, while workers are able to set their own hours.

A key challenge in the resource allocation problem is to find near-optimal policies to serve different customers with random demands/revenues, using a fixed pool of capacity (properly configured). In this paper, we study the properties of three classes of allocation policies—responsive (with perfect hindsight), adaptive (with information updates), and anticipative (with forecast information) policies. These policies differ in how the information on actual demand and revenue of each customer is being revealed and integrated into the allocation decisions. We show that the analysis of these policies can be unified through the notion of “persistency” (or service level) values—the probability that a customer is being (completely) served in the optimal responsive policy. We analyze and compare the performances of these policies for both capacity minimization (with given persistency targets) and revenue maximization (with given capacity) models. In both models, the performance gaps between optimal anticipative policies and adaptive policies are shown to be bounded when the demand and revenue of each item are independently generated. In contrast, the gaps between the optimal adaptive policies and responsive policies can be arbitrarily large. More importantly, we show that the techniques developed, and the persistency values obtained from the optimal responsive policies can be used to design good adaptive and anticipative policies for the other two variants of resource allocation problems. This provides a unified approach to the design and analysis of algorithms for these problems.

We propose a new multiple-benchmark tracking-error model for the portfolio selection problem. We construct a strategy to track the performance of the highest return from a set of benchmarks, by minimizing a quadratic loss function. The portfolio rule constructed this way retains features of the linear combination rule, with weights mainly determined by the probability that each benchmark portfolio attains the highest return among all the benchmark portfolios considered. We use this approach to deal with the issue of estimation errors in portfolio construction. A common strategy here is to use a re-sampling technique to construct portfolios from the data, and use the average of the portfolio weights as the solution. We can view each portfolio constructed from sampling as a benchmark, and use the tracking error method proposed in this paper to enhance the performance of any portfolio selection method. We perform comprehensive numerical experiments with various empirical data sets to demonstrate that our approach can consistently provide higher net Sharpe ratio (even after accounting for transaction cost), higher net aggregate return, smaller tracking error, and lower turnover rates in out-of-sample performance, compared to twelve dierent benchmark portfolios proposed in the literature, including the equally weighted portfolio (the 1=n strategy).