We introduce and study a fair assortment planning problem, where any two products with similar merits are offered similar visibility. We propose a framework to find near-optimal solutions to this problem, using the Ellipsoid method and a separation oracle to its dual. We then develop two approximate separation oracles, which result in a polynomial-time 1/2-approx. algorithm and an FPTAS for our problem. We conclude with a case study on a movie dataset, showing the efficacy of our algorithms.

Social planners often aim to equitably and efficiently ration a social good among agents whose (possibly correlated) demands realize sequentially. We design a simple adaptive policy that simultaneously achieves the best-possible guarantees on the expected minimum fill rate and the minimum expected fill rate, where each agent's fill rate is determined by an irrevocable, one-time allocation. We complement our results with a numerical study motivated by the rationing of COVID-19 medical supplies.

To solve the cold start problem on advertising platforms, we build a data-driven optimization model that captures the essential trade-off between short-term revenue and long-term market thickness on the platform, and propose a bandit algorithm to solve the problem. We also demonstrate the effectiveness of our algorithm via a novel two-sided randomized field experiment, and show our algorithm increases the cold start success rate by 62% and boosts the platform’s overall market thickness by 3.1%.

We present a dynamic experimental design for settings where the experimental units are coarse (e.g. to mitigate interference). `Region-split' experiments on online platforms are one such setting. Our design, dubbed Synthetically Controlled Thompson Sampling (SCTS), minimizes the cost (i.e. regret) associated with experimentation at no meaningful loss to inferential ability. We provide theoretical guarantees and experiments highlighting the merits of SCTS relative to fixed and switchback designs.

We use a text mining approach to extract waiting complaints from over 0.72 million online customer reviews of restaurants and conduct difference-in-differences regressions to estimate the impact of Virtual Queue (VQ). We find that VQ reduces customers' pre-process waiting complaints and does not lead to in-process waiting complaints increase. VQ also enhances customers' overall satisfaction. Service providers who face high substitutability or offer low-value service are benefited most from VQ.

When managing service systems, it is common to use priority rules based on some operational criteria. We consider the societal implications of such individual-focused priority policies, when individuals are members of broader population groups. We find that optimal service policies can lead to significant inequity across population groups. We propose policies that generate equitable outcomes across populations with little, or at times, even no additional system cost.

Users in online communities can ask questions and other users can answer these questions. Generally, question answerers get rewards while the askers gain knowledge if their questions get answered. We model the community as a stochastic game and find how users decide to participate in such communities. We theoretically validate the empirically observed network structure in such communities. Further, we find that the number of users in the community is non-monotonic in the participation cost.

Motivated by the debate around drivers' welfare in on-demand transportation, we propose a framework to evaluate current practices and possible alternatives. The platform allocates time slots to drivers, who are strategic agents maximizing their utility, which depends on their temporal preference (when to drive), slots they are allocated, and time they spend on-road. We use our framework to evaluate existing policies and propose improvements with respect to platform profit and drivers' wages.

We use 5-year data from a hospital with 16 inpatient units to empirically examine whether and how much congestion propagates through the network of inpatient units. We find that the magnitude of the congestion spillover is substantial in our study hospital. We then use counterfactual analyses to empirically identify the bottleneck unit---the unit that has the biggest impact on system performance when an intervention is applied to increase its capacity.

Via an app we developed, we identify a barrier to digital information access by the poor – data shortages. In a Mumbai slum, we randomly assigned respondents to a data plan with daily replenishment cycles – or a standard plan. Our data reveal that absent caps, respondents binge on YouTube and social media, resulting in subsequent data shortages. The capped plan increases late-plan access of WhatsApp invites to health camps, increases attendance at these camps, and reduces social media checking.

Emergency Department (ED) clients lack knowledge about the various elements their personal ED journey comprises. In a field study (N=18,903) we provide "Personalized Information about Operations and Time (PIOT)" – (frequently updated) information about the procedures and anticipated wait times for each patient. We show PIOT improves people’s attitudes, and PIO (operational-only information) reduces patients' abandonment. Providing PIOT offers a novel approach to improving healthcare service.

In this paper, we study how the boarding patients can affect the treatment time in emergency departments (EDs). First, we conduct an empirical analysis using a dataset from eight EDs and show that the relationship has an inverted U-shape. Next, we recognize two mechanisms behind it: (i)the additional workload on ED physicians, (ii) the hospitalist visits triggered by boarding congestion. Finally, we propose two interventions that can jointly mitigate this impact up to 68% in a tertiary hospital.

We introduce a MILP formulation for patient prioritization in the ED that incorporates a breakdown of predicted patient LOS. We propose an SAA based reformulation that can be solved efficiently and provide guarantees on its convergence, stability and sample complexity. A study of 40000 patient visits from a US hospital shows that our solution increases ED throughput by >50% and decreases average waiting time by >75%. Our method displays desirable properties of fairness in patient prioritization.

We present a queueing games framework to investigate how sharing of real-time queue-length information at emergency department (ED), where urgent patients receive priority over nonurgent patients, influences nonurgent patients' decision to enter or balk the ED queue, and how it affects the overall social welfare of patients who visit the ED. We show that under certain conditions, it may be better to partially reveal ED queue-length information rather than making ED queues completely transparent.

In a field experiment in an Emergency Department, we found that providing delay information improves patients' waiting satisfaction by 81%, and decreases their likelihood of reneging by 14%. The announced delay acts as a reference point against which the patients compare the actual delay. Following Prospect Theory, we found that patients are loss-averse that the likelihood of LWBS is much lower when they wait a shorter time than announced than when they wait a longer time than announced.

Standing orders allow triage nurses in EDs to order tests for certain medical conditions before the patient sees a physician, which could reduce the patient’s LOS. Medical literature documents the use of standing orders decreasing average ED LOS for the patient subject to standing orders. We model operational impact of standing orders and introduce a threshold based congestion-sensitive policy which performs well wrt overall average ED LOS across a wide range of scenarios.

Strategies for scheduling patients to minimize resource and waiting costs is a key problem in healthcare operation. The provider should decide patients to schedule as well as future appointment times to assign. We consider the multi-class advance scheduling problem with random arrivals. We propose and analyze the optimal policy that determines patients and appointments for the current state, effectively prioritizes the patients, and efficiently balances the workload across the booking window.

We propose a data-driven method for surgery scheduling that combines AI algorithms with stochastic optimization techniques. While existing approaches only consider the surgery type to differentiate between surgeries, our method allows us to consider different surgery-, patient- and physician-related aspects. We find that our method outperforms an existing method which only takes procedure types into account, resulting in higher operating room utilization and lower waiting times.

The simultaneous consideration of dynamic inter-day and intra-day scheduling decisions is an established theoretical and practical problem that has remained open due to its highly stochastic nature, complex structure, and the curse of dimensionality. We develop the first analytical optimization model and theoretical results addressing this joint problem within a computationally tractable optimization framework with theoretical performance guarantees.

Appointment scheduling under uncertainty encounters a fundamental trade-off between capacity utilization and waiting times. In contrast to traditional approaches, we maximize capacity utilization while limiting waiting times. We derive a robust mixed-integer linear model, prove NP-hardness for the general problem, and optimality of well-known scheduling rules for special cases. Using real patient data, we show that our approach is a win-win solution to this fundamental trade-off.

Using distributionally robust optimization (DRO) that accounts for service times’ correlation, we study a two-sequential-server appointment scheduling problem. We observe that the optimal schedule can lead to imbalanced waiting times in the two servers, concentrating on the downstream server. To rebalance the waiting times without rescheduling patients, we adopt an idea in the queueing literature to strategically idle (SI) the upstream server. A DRO model is used to find the optimal SI policy.

We present a personalized and probabilistic framework to model the risk of hospital admissions for patients (with and without COVID-19) that attended an Emergency Department (ED) during the pandemic. As predictors, we use patient demographics, measures of ED crowdedness, and the triage information, and investigate if population-level data (such as human mobility, number of COVID-19 cases, vaccination status etc.) could help improve the prediction accuracy of admission risk at the patient level.

We consider the problem of designing a sequential clinical trial with a fixed budget in order to find the best treatment as a function of predictive and prognostic patient covariates. We propose computationally tractable heuristics based on the expected value of information that perform well and are asymptotically optimal in the limit of large sample size. We show the benefits of incorporating predictive and prognostic covariates in allocation policies for learning the best treatment strategy.

We exmaine potential learning at the manufacturing facility level. Considering drug shortages as a manufacturing disruption, we investigate the two sources of learning: (1) experience of recovery from disruptions in the past and (2) experience of recovery from on-going disruptions. In addition to these learning effects, we also examine whether the two learning sources are substitutes or complements, and how the diversity of disruption resolution experience moderate these learning effects.

Many US states have adopted telehealth pay-parity policies requiring payers to reimburse healthcare providers equally for telehealth and office visits. But telehealth may require a duplicate visit for a physical exam. We analyze a three-stage game to study the impact of telehealth reimbursement on provider's operational decisions, where patients choose strategically between telehealth and office channels. We find that pay parity can decrease patient access and discuss its policy implications.

Current clinical guidelines for hypertension provide physicians with general suggestions for first-line pharmacologic treatment, but do not consider patient-specific characteristics. We utilize electronic health record data to develop personalized predictions and prescription models for hypertensive patients. We demonstrate a 15.87% improvement over the standard of care and propose a novel interactive dashboard to facilitate the deployment of the derived models in the clinical practice.

Rapid and accurate detection of local outbreaks is critical to tackle resurgent waves of COVID-19. A fundamental challenge in case growth rate estimation, a key epidemiological parameter, is balancing the accuracy vs. speed tradeoff for small sample sizes of counties. We present “Transfer Learning for Generalized Random Forests” (TLGRF), a novel framework which uses relevant features affecting the disease spread across time and counties to obtain more robust and timelier county-level estimates.

We address a set of problems that occur in the development of liquid biopsies via the lens of active sequential hypothesis testing (ASHT). Motivated by applications in which the number of hypotheses or actions is massive, we propose efficient algorithms and provide the first approximation guarantees for ASHT, under two types of adaptivity. We numerically evaluate the performance of our algorithms using both synthetic and real-world DNA mutation data.

We enhance the multi-armed bandit model by considering endogenously non-stationary rewards – specifically rewards that are parametric functions of policy histories (learning). We further incorporate queueing costs, fairness, and arm correlation. We propose the L-UCB, FL-UCB, and QFL-UCB algorithms to solve our model, prove its logarithmic regret, and apply it to split-liver transplantation.

We adapt online learning techniques to scenarios with time-to-event data, where there is a delay between choosing an arm and observing feedback that is endogenous to the quality of the arm. We posit a multi-armed bandit algorithm with a cox-proportional hazards estimator, prove guarantees on the regret under this algorithm, and analyze its performance on a dataset of metastatic breast cancer clinical trials, comparing it to that of other adaptive allocation schemes.

Lack of treatment adherence significant barrier to reducing the global disease burden of tuberculosis (TB). We study the design of targeted interventions for a treatment adherence support platform running in Kenya, whose goal is to help patients on TB treatment. We show empirically that there is large heterogeneity in treatment effects of interventions, and we devise a novel online learning policy based on Thompson Sampling that significantly outperforms the currently employed policy.

This paper studies a multi-channel healthcare system with both a telemedicine and an in-person channel. Using a DID identification strategy, we find that introducing multiple channels has significant impact on the in-person channel and system level, including an increase in case complexity and planned consultation time by 20% for in-person consultations. In addition, we observe an increase in system capacity, a reduction in wait time for in-person appointments by 37.5%.

In this paper, we provide an empirical investigation into the role of horizontal and vertical integration of orthopaedic surgeons in driving heterogeneity in the impact of the Comprehensive Joint Replacement (CJR) program. Using a difference-in-differences approach, we find hospitals with high horizontal and vertical integration see an increase in both hospital costs and complication rates of 3.17% 1.17, respectively, while others see either a decrease or no change in these measures.

Comorbid depression could lead to a 100% increase in the cost of care for diabetes. Clinical trials have demonstrated that depression care through care managers in a primary care setting (called Collaborative Care) leads to faster depression remission. We present a mathematical modeling approach that determines the optimal allocation of care managers' time to enrolled patients towards improving clinic revenue and patient health outcomes.