Predictive Analytics, Business Modeling and Optimization in Healthcare Operations Management

Risk aversion is natural when planning for healthcare services. Surgery planning and assigning example will be used to illustrate co-advancement in chance-constrained optimization research to allow consideration of risk-aversion in operational planning problems. We will show the implications of robustness towards out-of-sample risk aversion, and problem solvability when a limited number of data samples are available. We will also see how the classical cut-generation theory is extensible to the more general settings of robust chance constraints.

Geographical mismatch between demand for care and availability of healthcare resources has been a major challenge during the COVID-19 pandemic. As such, inter-hospital patient transfers emerged as a key aspect of the pandemic response in many countries, including the US and Canada. We propose and study a class of transient queueing control problems to gain insights into the structure of “good” transfer policies. Specifically, we consider multiple parallel queues, starting with a large and possibly unbalanced initial condition, and seek discrete-time dynamic transfer policies that minimize the expected total cost over a finite-horizon or until the system reaches a desirable state. Costs are incurred as customers wait in queues, and are transferred between queues – including both variable and fixed transfer costs. We characterize the structure of the optimal policy for an associated discrete-time fluid control problem and investigate the robustness of the structure for the original stochastic problem. Using simulation experiments, we illustrate the performance of fluid-based policies and examine the value of optimal transfers. Finally, we discuss how the results can facilitate the search for an optimal transfer policy in a high-fidelity stochastic model of patient flow.

Ryan Buell, Kamalini Ramdas, Nazli Sonmez, Kavitha Srinivasan, Rengaraj Venkatesh

Unlike promising new drugs or devices, high-potential operational innovations in care delivery are often slow to spread, due in part to difficulties in evaluating their impact. We scientifically examine the impact of a little-known yet potentially high-impact and scalable care delivery innovation. In a shared medical appointments (SMA), a group of patients with similar chronic conditions meets with their doctor at once, and each in turn receives one-one care, albeit not in private.  In a 1,000-patient, multi-stage, randomized controlled trial of shared medical appointments vs. traditional one-on-one appointments for regular glaucoma follow up at the Aravind Eye Hospital in Pondicherry, India, we  find that SMAs increase patients’ satisfaction, knowledge and healthy behaviors. A key barrier to wider adoption of SMAs is the concern – raised by both patients and doctors – that patients may engage less due to lack of privacy. We use verbal and nonverbal engagement measures coded from 20,000 minutes of video data covering every  trial appointment to throw light on how interactions in the SMAs arm of our trial differ from those in one-on-one appointments arm, and to gain insight into how these interactions predict healthy behaviors. We will discuss how our approach could be used in the context of other care delivery innovations.

Problem definition: Healthcare systems in resource-limited settings rely on diagnostic networks in which medical samples (e.g., blood, sputum) and results need to be transported between geographically dispersed healthcare facilities and centralized laboratories.

Academic/practical relevance: Existing sample transportation (ST) systems typically operate fixed schedules, which do not account for demand variability and lead to unnecessary transportation visits as well as delays.

Methodology: We design an optimized sample transportation (OST) system that comprises two components: (i) a new approach for timely collection of information on transportation demand (samples and results) using low-cost technology based on feature phones, and (ii) an optimization-based solution approach to the problem of routing and scheduling courier trips in a multistage transportation system.

Results: Our solution approach performs well in a range of numerical experiments. Furthermore, we implement OST in collaboration with Riders For Health, who operate the national ST system in Malawi. Based on analysis of field data describing over 20,000 samples and results transported during July–October 2019, we show that the implementation of OST routes reduced average ST delays in three districts of Malawi by approximately 25%. In addition, the proportion of unnecessary trips by ST couriers decreased by 55%.

Managerial implications: Our approach for improving ST operations is feasible and effective in Malawi and can be applied to other resource-limited settings, particularly in sub-Saharan Africa.

This talk focuses on the design and analysis of experiments conducted on a set of units, such as clinics or geographical regions, over multiple periods where the starting time of the treatment may vary by unit. We aim to minimize the required sample size by using an adaptive experimental design, which involves updating treatment assignments and determining whether to continue the experiment based on data collected in previous periods. We propose an algorithm that addresses the challenges at both the design stage and estimating the treatment effects, ensuring valid post-experiment inference that accounts for the adaptive nature of the design. Our approach utilizes dynamic programming to optimize the speed of treatment rollout in subsequent periods, and employs sample splitting to ensure that the obtained estimates are consistent and asymptotically converge to a normal distribution. An appealing property of the proposed scheme is that the final treatment effect estimation uses all of the data, incurring no efficiency loss compared to an oracle who would have access to the same design at the beginning of the experiment. Using a large national database of influenza patients, we demonstrate that our proposed approach can reduce the opportunity cost of testing the effectiveness of public health interventions by over 50%, compared to the best non-adaptive design benchmarks.

Mass vaccination is essential for pandemic control, but long queues can increase infection risk. We study how to schedule arrivals at a mass vaccination site to minimize a tri-objective function of a) expected number of infections acquired while waiting, b) throughput, and c) overtime. Leveraging multi-modularity results of a related optimization problem, we construct a solution algorithm and compare our results to an equally-distributed, equally-spaced schedule. While we find that the latter sits near the pareto-optimal frontier, it is located away from a sharp elbow in the tradeoff between infections and overtime. Specifically, the elbow-policy achieves approximately 55% fewer expected infections for nearly the same expected overtime.  We also discuss managerial insights around the structure of the optimal schedule and compare it to the well-known “dome-shaped” policies found in other appointment scheduling contexts.

There is an increasing amount of evidence that Machine Learning and Artificial Intelligent algorithms can be used to enhance clinical care. In this talk, I address two critical aspects that can significantly improve the impact of such algorithms in healthcare practices: (1) moving beyond associations, and creating algorithms capable of causal inference under ambiguity, and (2) a human-algorithm “centaur” model of care and decision-making, in which the power of human intuition is combined with the outstanding capabilities of algorithms. I describe our latest research at the Public Impact Analytics Science Lab (PIAS-Lab) at Harvard on these subjects, and discuss findings based on our various collaborations with the Mayo Clinic as well as some other public and private organizations.

Practitioners and academics have long appreciated the benefits of covariate balancing in designing lab and field experiments. However, covariate balancing remains challenging when experimental units arrive sequentially. In this paper, we study an online experimental design problem, which we refer to as the “Online Blocking Problem.” In this problem, experimental units with heterogeneous covariate information arrive sequentially and must be immediately assigned into either the control or the treated group, with an objective of minimizing the total discrepancy, which is defined as the minimum weight perfect matching between the two groups. To solve this problem, we propose an experimental design approach, which we refer to as the “Pigeonhole Design.” The pigeonhole design first partitions the covariate space into smaller spaces, which we refer to as pigeonholes, and then, when the experimental units arrive at each pigeonhole, balances the number of control and treatment units for each pigeonhole. We analyze the theoretical performance of the pigeonhole design and show its effectiveness by comparing against two well-known benchmark designs: the match-pair design and the completely randomized design. This is joint work with Zijie Zhou (MIT).

Negative side effects from taking a drug, termed adverse drug reactions (ADRs), cause thousands of deaths a year in the U.S. alone. After drug approval, regulators such as the U.S. Food and Drug Administration (FDA) rely on post-approval surveillance from observational data sources such as spontaneous adverse event reporting systems to detect potential safety issues. These systems contain large volumes of safety reports voluntarily submitted by patients and their doctors, and regulators generate hypotheses about potential safety issues (termed safety signals) by identifying side effects that occur at a disproportionately high rate in patients taking a drug versus patients taking other drugs for the same condition. Reliance on biased observational data — due to selection and reporting differences among patients — can result in regulators flagging safety issues that are not truly present or in missing real safety issues. In this work, we seek to enhance the hypothesis generation step of safety signaling via a Bayesian methodology that combines pre-approval clinical trials and post-approval observational data for multiple ADRs. We use data from more common adverse events to quantify the direction and magnitude of bias in observational data as compared to clinical trial data and use it to debias the observational data for more rare adverse events. Via analytical modeling, we characterize the situations in which this methodology may be particularly valuable to regulators compared to the status quo. A detailed numerical evaluation using real-world clinical trials and post-approval observational data suggests that our approach can in some cases significantly reduce the rate and cost of errors committed in safety signaling.

Machine learning has the potential to increase the scale of real-world data curated from electronic health records, but maintaining a high standard of data quality is important to avoid biasing downstream analyses. To increase scale without compromising quality, we propose a hybrid data curation methodology that employs both manual abstraction by clinical experts and automated extraction by machine learning models. Our methodology makes the determination about when to employ manual abstraction using a confidence score associated with each model output. We describe a process for selecting confidence thresholds based on simulations validated against a reference-standard labeled dataset. To establish the fitness of our methodology for retrospective research, we apply it to a multi-variable cohort selection task on a large real-world oncology database.