This article was downloaded by: [137.73.144.

138] On: 30 August 2019, At: 05:35

Publisher: Institute for Operations Research and the Management Sciences (INFORMS)
INFORMS is located in Maryland, USA

Management Science
Publication details, including instructions for authors and subscription information:
http://pubsonline.informs.org

Data-Driven Patient Scheduling in Emergency

Departments: A Hybrid Robust-Stochastic Approach
Shuangchi He, Melvyn Sim, Meilin Zhang

To cite this article:

Shuangchi He, Melvyn Sim, Meilin Zhang (2019) Data-Driven Patient Scheduling in Emergency Departments: A Hybrid Robust-
Stochastic Approach. Management Science

Published online in Articles in Advance 01 May 2019

. https://doi.org/10.1287/mnsc.2018.3145

Full terms and conditions of use: https://pubsonline.informs.org/page/terms-and-conditions

This article may be used only for the purposes of research, teaching, and/or private study. Commercial use
or systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisher
approval, unless otherwise noted. For more information, contact [email protected].

The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitness
for a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, or
inclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, or
support of claims made of that product, publication, or service.

Copyright © 2019, INFORMS

Please scroll down for article—it is on subsequent pages

INFORMS is the largest professional society in the world for professionals in the fields of operations research, management
science, and analytics.
For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org
 MANAGEMENT SCIENCE
Articles in Advance, pp. 1–18
http://pubsonline.informs.org/journal/mnsc/ ISSN 0025-1909 (print), ISSN 1526-5501 (online)

Data-Driven Patient Scheduling in Emergency Departments:

A Hybrid Robust-Stochastic Approach
Shuangchi He,a Melvyn Sim,b Meilin Zhangc
a
Department of Industrial Systems Engineering and Management, National University of Singapore, Singapore 117576; b Department of
Analytics and Operations, NUS Business School, National University of Singapore, Singapore 119245; c School of Business, Singapore
University of Social Sciences, Singapore 599494
Contact: [email protected], http://orcid.org/0000-0003-4107-3946 (SH); [email protected],
http://orcid.org/0000-0001-9798-2482 (MS); [email protected], http://orcid.org/0000-0003-2880-8223 (MZ)

Received: November 22, 2015 Abstract. Emergency care necessitates adequate and timely treatment, which has un-
Revised: February 19, 2017; April 16, 2018 fortunately been compromised by crowding in many emergency departments (EDs). To
Accepted: May 24, 2018 address this issue, we study patient scheduling in EDs so that mandatory targets imposed
Published Online in Articles in Advance: on each patient’s door-to-provider time and length of stay can be collectively met with the
May 1, 2019 largest probability. Exploiting patient flow data from the ED, we propose a hybrid robust-
https://doi.org/10.1287/mnsc.2018.3145 stochastic approach to formulating the patient scheduling problem, which allows for
practical features, such as a time-varying patient arrival process, general consultation time
Copyright: © 2019 INFORMS distributions, and multiple heterogeneous physicians. In contrast to the conventional
formulation of maximizing the joint probability of target attainment, which is compu-
tationally excruciating, the hybrid approach provides a computationally amiable for-
mulation that yields satisfactory solutions to the patient scheduling problem. This
formulation enables us to develop a dynamic scheduling algorithm for making recom-
mendations about the next patient to be seen by each available physician. In numerical
experiments, the proposed hybrid approach outperforms both the sample average ap-
proximation method and an asymptotically optimal scheduling policy.

History: Accepted by Yinyu Ye, optimization.

Funding: The work of S. He was supported in part by the Singapore Ministry of Education Academic
Research Fund [Grant MOE2017-T2-1-012]. The work of M. Sim was supported in part by the
Singapore Ministry of Education Social Science Research [Thematic Grant MOE2016-SSRTG-059].
Supplemental Material: The e-companion is available at https://doi.org/10.1287/mnsc.2018.3145.

Keywords: healthcare operations • patient scheduling • robust optimization • stochastic programming • mixed integer programming •
queueing network

1. Introduction the Centers for Medicare and Medicaid Services (CMS)

Emergency department (ED) crowding and the conse- in the United States publishes quality measures of timely
quential delays have been a worldwide issue and re- emergency care for over 4,000 hospitals on their Hospital
ceived considerable attention from governments, Compare website; some of these measures are included
public media, and academic communities. ED crowding in the pay-for-performance program of the CMS. To
compromises the quality of and access to emergency address the crowding issue, governments and regula-
care, putting patients at great risk of treatment errors. tory organizations may set mandatory targets for
Numerous studies have revealed an association be- emergency care. In 2005, England’s National Health
tween crowding and increased morbidity and mortality Service mandated that 98% of ED patients must be
in EDs (McHugh 2013). For hospitals, ED crowding treated and either discharged home or admitted to an
damages their public reputation and incurs revenue loss inpatient ward within four hours of arrival. The
because of ambulance diversion and patients’ leaving implementation of this “four-hour rule” greatly im-
without being seen. As pointed out by Rabin et al. proved the percentage of patients spending less than
(2012), widespread crowding also impedes hospitals’ four hours in EDs from 77.3% in 2002–2003 to 97.2% in
ability to achieve national safety and quality goals, 2008–2009 (Weber et al. 2011).
compromises the healthcare system, and limits the re- Hoot and Aronsky (2008) summarized the common
gional capacity for disaster response. In many countries, causes of ED crowding, including an increasing de-
the performance of hospitals’ emergency care is closely mand for emergency care, insufficient hospital bed ca-
monitored by government agencies, and some key in- pacity, operational inefficiencies, and so forth. Effective
dicators are made public on a regular basis. For example, patient flow management is expected to be the solution
1
 He, Sim, and Zhang: Data-Driven Patient Scheduling in Emergency Departments
2 Management Science, Articles in Advance, pp. 1–18, © 2019 INFORMS

to excessive patient delays without direct capacity be seen whenever a physician becomes available. To
expansion. To evaluate the timeliness and efficiency of make timely recommendations, the scheduling algo-
emergency care, the National Quality Forum has en- rithm must be sufficiently efficient.
dorsed length of stay, door-to-provider time (i.e., the One may formulate an optimization problem to
time that a patient spends in the ED before being seen obtain the schedule by maximizing the joint probability
by a healthcare provider), and leaving without being of all waiting patients meeting the delay targets (please
seen as quality metrics (Welch et al. 2011). Because the see formulations (6) and (7) and related discussion).
percentage of leaving without being seen is closely re- With the joint probability of target attainment being
lated to patients’ door-to-provider times, we regard the the objective, such an optimization problem was first
two time metrics as major performance concerns. In studied by Charnes and Cooper (1963), who termed
general, door-to-provider times should be kept below this formulation the P model. The P model formulation,
certain safety limits according to each patient’s clinical however, is not widely used in practice, in part because
urgency. The widely used Emergency Severity Index, for evaluating the joint probability demands integra-
example, categorizes ED patients into five groups based tion in high dimensions, which is generally compu-
on their acuity levels and required medical resources; the tationally intractable, let alone solving the associated
recommended door-to-provider time targets range from nonconvex optimization problem.
“immediately” for resuscitation patients to “within one To tackle this issue, we incorporate features from
to two hours” for less urgent patients (Gilboy et al. 2011). robust optimization into our formulation by consid-
Both clinical and operational requirements impose strict ering a family of uncertainty sets. Associated with a given
time constraints on patient flow management. schedule, each uncertainty set in the family consists of
The focus of this paper is patient scheduling in EDs. the feasible consultation times that patients can take
From a modeling perspective, an ED can be viewed as without violating the mandatory delay targets. Unlike
a queueing network, with medical units being the conventional robust optimization formulations, where
nodes; patients being the customers; and beds, medical uncertainty sets are fixed, the hybrid approach searches
staff, and equipment being the servers (Armony et al. in the family for the uncertain set that has the largest
2015). Aside from prioritized customers and time- probability of all of its consultation times being feasible.
sensitive service requirements, this network is char- The schedule associated with the obtained uncertainty
acterized by frequent returning routes of customers set is the optimal solution to the hybrid formulation.
(i.e., after their initial consultations with a physician, For computational reasons, we restrict the family of
most patients undergo medical tests and return to the uncertainty sets to a collection of hyperrectangles.
same physician before eventually being discharged or Then, under the independence assumption of consul-
hospitalized). Although emergency physicians are re- tation times, the joint probability of all waiting patients
quired to provide treatment for a broad spectrum of meeting the delay targets is simply the product of
illnesses and injuries, their expertise and work rates the marginal probabilities of each individual patient
differ from one another. In other words, the servers of meeting his own delay target. In this case, computing
this network are heterogeneous. When there are mul- the joint probability does not involve high-dimensional
tiple patients waiting to be seen, their respective integration, which may greatly improve the compu-
physicians and the sequence of their consultations must tational efficiency of the scheduling algorithm. In nu-
be carefully scheduled to meet the stringent door-to- merical experiments, the hybrid robust-stochastic
provider and length of stay targets. However, the approach outperforms both the sample average ap-
aforementioned features, including the complex net- proximation (SAA) method and an asymptotically
work structure, server heterogeneity, highly uncertain optimal policy; more details are in Section 7.
patient arrival processes, and time-sensitive service The hybrid robust-stochastic approach is of both
requirements, all pose challenges in solving the patient practical and methodological importance. First, al-
scheduling problem. though the hybrid formulation is essentially a mixed
To address this problem in a practical setting, we integer program, solving this problem is practically
propose a hybrid robust-stochastic approach to exploiting efficient and allows for real-time scheduling in EDs. As
patient flow data for real-time patient scheduling. Our a dynamic approach driven by data, it allows for
intention is to maximize the percentage of patients practical features, such as a time-varying patient arrival
whose door-to-provider times and lengths of stay are process, general consultation time distributions, and
within the mandatory targets. Because the patient ar- heterogeneous physicians. In the literature on schedul-
rival pattern is highly variable, we refrain from making ing of queueing networks, these features are generally
assumptions, such as the arrival rate and the inter- absent from existing network models. As a result, the
arrival time distribution, about future patient ar- existing scheduling policies may not perform as well in
rivals. Using the data of existing patients, the dynamic practice. Second, the hybrid formulation represents an
scheduling algorithm will determine the next patient to alternative perspective on solving the P model problem,
He, Sim, and Zhang: Data-Driven Patient Scheduling in Emergency Departments
Management Science, Articles in Advance, pp. 1–18, © 2019 INFORMS 3

the objective of which is to maximize the feasibility With the feature that patients may return to the same
probability of a set of randomly perturbed linear con- physician several times, a refined Erlang-R model for
straints. Conceivably, the hybrid formulation may pro- the occupancy process was proposed by Yom-Tov and
duce near-optimal solutions at a far lower computa- Mandelbaum (2014). Saghafian et al. (2012) analyzed
tional expense. As illustrated by numerical examples in the practice of patient streaming (i.e., separating pa-
Section 7.2, our approach may provide a highly efficient tients based on the predictions of whether they will be
alternative to the SAA method. In addition to patient discharged or hospitalized) in EDs and proposed an
scheduling, similar problems arise from other stochas- improved streaming scheme. Saghafian et al. (2014) pro-
tic systems with time-sensitive service requirements; posed a new triage system based on both clinical ur-
Section 8 has more discussion. gency and treatment complexity for improving patient
The remainder of this paper is organized as follows. safety and operational efficiency.
The related literature is reviewed in Section 2. We in- Patient scheduling in EDs was studied by Huang
troduce the queueing network model for EDs in Section 3. et al. (2015), whose work is the most relevant to ours
In Section 4, we present a tractable approach to in the literature. In their paper, the ED is modeled as
solving the patient scheduling problem based on a multiclass queueing network with service deadlines
a hybrid robust-stochastic formulation. This hybrid and feedback routes. The authors proposed a simple,
formulation is translated into a mixed integer program yet highly effective scheduling policy that is capable
in Section 5. By introducing additional delay constraints, of striking a balance between maintaining acceptable
the hybrid formulation is incorporated into a dynamic door-to-provider times and mitigating congestion. By
scheduling framework in Section 6, which enables us means of heavy traffic analysis, they proved that, under
to solve the patient scheduling problem sequentially a simplified setting, their proposed scheduling policy is
according to a stochastic patient arrival process. We asymptotically optimal for reducing the total conges-
provide a comprehensive data-based numerical study in tion cost subject to constraints on door-to-provider
Section 7, where the hybrid approach is compared with times. This scheduling policy serves as an important
both the SAA method and an asymptotically optimal benchmark for our hybrid approach; Section 7.3 has
scheduling policy. The paper is concluded in Section 8, a comparison between these two approaches.
where some potential applications and future research Although the aforementioned queueing models are
are discussed. We leave the construction of nonanticipative able to represent basic operational characteristics of an
arrangements, all proofs, and additional simulation ED, they may be overly simplistic and incapable of
results to the e-companion. capturing some salient features. For a queueing model to
Let us close this section with frequently used nota- be analytically tractable, one may require probabilistic
tion. Scalars and vectors are denoted by lowercase and assumptions, such as exponential interarrival and ser-
bold letters, respectively. Calligraphic letters are used vice time distributions, stationary arrival processes, and
for sets, such as (, and we use |(| for the cardinality of homogeneous servers. As pointed out by Bertsimas et al.
the set. Random variables and vectors are denoted with (2011), the performance analysis of queueing networks
a tilde mark, such as s̃ and s̃. We assume that all ran- is largely unsolvable without these assumptions. How-
dom variables and vectors are defined on a common ever, as the ED environment is complex and changes
probability space, where P(A) is the probability that frequently, such assumptions may not be appropriate.
event A occurs. We reserve E(s̃) for the expectation of Conceivably, control policies obtained under these as-
a random variable s̃. sumptions may not necessarily work well in practice. In
contrast, the proposed hybrid formulation does not rely
2. Related Literature on such assumptions. We would thus expect this data-
We sketch relevant studies to position our work within driven approach to better fit the ED environment.
the literature. Both the literature on patient flow man- Even if the queueing network model is analytically
agement and the literature on optimization of queueing tractable, it is still difficult to find an optimal dynamic
networks are extensive and well established. It is not control policy with delay or throughput time con-
our intention to be exhaustive. straints. Most studies in the literature focus on simple
For analysis and control purposes, EDs are usually policies that can be proved optimal in some asymptotic
modeled as queueing networks. Although most stud- sense (e.g., Doytchinov et al. 2001, Plambeck et al. 2001,
ies are simulation based (see Connelly and Bair 2004, Maglaras and Van Mieghem 2005, and Huang et al.
Sinreich and Marmor 2005, and the references therein), 2015). Some aforementioned simplistic assumptions,
several simplified queueing models are used in analytical along with a heavy traffic condition, are necessary for
studies. For example, to determine the staffing level of asymptotic optimality to hold. The performance of
physicians, the ED occupancy process is described by these policies may be suboptimal when the simplis-
a time-varying Erlang-C model in Green et al. (2006) and tic assumptions are not satisfied. Moreover, the con-
by a time-varying Erlang-B model in de Bruin et al. (2010). trol actions of these policies depend on service time
 He, Sim, and Zhang: Data-Driven Patient Scheduling in Emergency Departments
4 Management Science, Articles in Advance, pp. 1–18, © 2019 INFORMS

distributions only through their first moments. For naturally from the formulation of reserve provision
these policies to be near optimal, deadlines for delay or in their paper, whereas in our approach, hyper-
throughput times must be on a higher order of mag- rectangular sets serve primarily as a heuristic proxy
nitude than service times, which may not be a rea- for feasible sets in solving the P model problem. As-
sonable assumption in the ED setting. In contrast, the sociated with a joint probability measure, these
hybrid robust-stochastic approach allows for multiple hyperrectangular sets should not be understood as
heterogeneous servers, time-varying arrival processes, uncertainty sets in the usual sense. From the meth-
and arbitrary traffic conditions. Distributional infor- odological perspective, the main theme of their paper is
mation obtained from patient flow data is fully used in how to confine both uncertainty sets and admissible
constructing uncertainty sets. In other words, the hy- policies to affine structures so that the resulting for-
brid approach can make use of entire service time mulation becomes a convex optimization problem. In
distributions, which turns out to be a considerable ad- contrast, we convert the patient scheduling problem
vantage over the previous scheduling policies. In nu- into a tractable mixed integer program, relying on the
merical experiments in Section 7.3, the hybrid approach fact that, within each hyperrectangle, the worst case
outperforms the asymptotically optimal scheduling only occurs at a single boundary point (discussed in
policy proposed by Huang et al. (2015), even though Theorem 1 in Section 4). It is also worth mentioning
the ED is in heavy traffic. that, as decision variables, uncertainty sets in both
Under the framework of robust optimization, the studies are subject to joint constraints with control
performance analysis of queueing networks was policies. In other words, the families of adjustable
studied by Bertsimas et al. (2011), Bandi and Bertsimas uncertainty sets may change with specific policies. This
(2012), and Bandi et al. (2015). In their papers, ran- is similar to the robust optimization formulation pro-
domness in arrival and service times is modeled by posed by Spacey et al. (2012), who studied software
polyhedral uncertainty sets using limit laws in prob- partitioning with multiple instantiation (i.e., assigning
ability theory. More specifically, the law of the iterated code segments of a computer program to multiple
logarithm was considered by Bertsimas et al. (2011), execution locations so as to minimize the overall
and the generalized central limit theorem was con- program run time). The only decision variable in their
sidered by Bandi and Bertsimas (2012) and Bandi et al. problem is the software partition, which also deter-
(2015) for constructing uncertainty sets. Using this mines the uncertainty set of location-aware control
robust optimization approach, the authors obtained flows. Unlike in our study, their uncertainty set is not
performance bounds for queueing networks. Although adjustable, and therefore, it is not a decision variable
our approach is also inspired by robust optimization, of the optimization problem.
it stems from a completely different perspective. As
opposed to conventional robust optimization formu- 3. The Controlled Queueing
lations where uncertainty sets are specified as fixed Network Model
constraints, our approach investigates a family of un- We use a queueing network to model the ED, which is
certainty sets and searches for the schedule that controlled by a centralized patient scheduling system
“maximizes” the uncertainty set within the family. In to meet requirements for door-to-provider times and
this sense, the obtained schedule is the most “robust” lengths of stay. Based on the state of current patients, the
solution to the patient scheduling problem. scheduling system will make sequential recommenda-
Our hybrid robust-stochastic approach shares some tions for the next patient to be seen for each physician.
features with a concurrent, independent study by The general flow of patients goes through the ED
Zhang et al. (2017), who considered robust optimal according to the following process. Patients arrive
control of constrained linear systems with adjustable at the ED in a stochastic and nonstationary manner.
uncertainty sets. The formulation of their problem is After registration, they will be triaged by a nurse and
motivated by reserve provision in electrical grids, assigned to several urgency groups based on their
where the reserve capacity of power is required to be acuity levels and other concerns. The door-to-provider
periodically adjusted for cost saving without sacrificing times of patients in each urgency group should be kept
necessary power consumption. In their formulation, below a prescribed safety limit, and the safety limits of
a series of adjustable uncertainty sets is used to rep- the urgency groups may differ from one another. Then,
resent the reserve capacity, thus becoming decision patients will stay in a waiting area until they are called
variables as in our approach. Zhang et al. (2017) also to be seen by a physician. These patients will be re-
investigated uncertainty sets of special geometric forms ferred to as new patients. After initial consultations,
to render their optimization problem tractable. Despite some patients may leave the ED, whereas others may
these analogous features, our study differs significantly undergo diagnostic tests, such as x-rays and blood
from their work in the following aspects. From the tests, or receive treatment from a nurse. When the test
modeling perspective, adjustable uncertainty sets arise result is ready or the treatment is completed, the patient
He, Sim, and Zhang: Data-Driven Patient Scheduling in Emergency Departments
Management Science, Articles in Advance, pp. 1–18, © 2019 INFORMS 5

will return to the waiting area and become a returning Figure 1. Queueing Network Model of Patient Flow in an ED
patient, waiting to be examined by the same physician.
A patient may see the same physician several times
before eventually being discharged or hospitalized.
The scheduling system will determine the assign-
ment of patients to each physician and the order of their
consultations. A new patient can be assigned to any
available physician, whereas a returning patient must
be seen by the physician with whom he consulted
initially. We assume that the scheduling system does
not manage patients who are waiting for tests or
treatment by a nurse, because those patients are typi-
cally served on the first-come, first-served (FCFS)
basis. A patient scheduling system is critical for
the mitigation of ED crowding, because simple prior-
itization rules are incapable of balancing door-
to-provider times and lengths of stay. If physicians
give priority to new patients to reduce their door-to-
provider times, returning patients have to spend more
time waiting and form a long queue. By Little’s law, the
mean length of stay will be prolonged. In particular,
patients who need multiple consultations will have For i ∈ ( and j ∈ ), let s̃ij be the consultation time of
long total waiting times, which may create a long tail in patient i if he would be seen by physician j. For i ∈ #j , s̃ij
the distribution of lengths of stay. When the ED be- is interpreted as the remaining consultation time of
comes crowded, the lengths of stay of these patients patient i as he is being seen by physician j. We assume
will be likely to exceed the mandatory target. Giving that {s̃ij : i ∈ (, j ∈ )} is a set of mutually independent
priority to returning patients can effectively shorten the random variables and use Fij to denote the cumulative
queue length in the waiting area, thus reducing lengths distribution function of s̃ij . Because the physicians
of stay. This strategy, however, will inevitably prolong may be heterogeneous, even if i ∈ ( is fixed, Fij may
the door-to-provider times of new patients, putting vary for different j ∈ ). Each Fij can be estimated using
them at risk of treatment delays. Moreover, to maintain the records of physician j’s consultation times and may
operational efficiency, the expertise of each physician depend on the physician’s expertise as well as the
must be considered in deciding the next patient to be patient’s status (new or returning), triage information,
seen. In general, it would be difficult to find a rule of preliminary diagnosis, etc. In our implementation, Fij is
thumb for patient scheduling under constraints on both taken to be the empirical distribution function of
door-to-provider times and lengths of stay. a selected sample of consultation times. Therefore, we
The controlled queueing network model is depicted assume each s̃ij to be a discrete random variable, with
in Figure 1. The scheduling system determines the next values taken from a finite set of positive numbers
patient to be seen when a physician completes a con- 6ij  {sij (1), . . . , sij (Nij )}. We use sij and s̄ij to denote the
sultation, and determines a new patient's physician smallest and greatest numbers, respectively, in 6ij .
when he comes to the waiting area finding at least one Let s̃  (s̃ij )i∈(,j∈) be the random vector of all of these
free physician. Let t be such a time, and consider the ED consultation times. Then, s̃ takes values from the product
at this moment. Let ) be the set of physicians, 1 be the space 6  ∏i∈(,j∈) 6ij .
set of new patients, # be the set of patients being seen, The assignment of waiting patients to physicians is
and 5 be the set of returning patients, where the de- specified by a function ϕ : 0 → ), where ϕ(i) is the
pendence on t is suppressed for notational conve- physician of patient i. Because returning patients must
nience. For j ∈ ), we use #j to denote the set of patients be seen by their initial physicians, the assignment should
being seen by physician j and 5j to denote the set of satisfy
returning patients to be seen by physician j. Then, ϕ(i)  j for i ∈ 5j and j ∈ ). (1)
#  ⋃j∈) #j and 5  ⋃ j∈) 5j . Moreover, #j  Ø if and
only if physician j is available at time t; otherwise, #j Sequencing decisions are specified by a correspon-
has exactly one patient. Let 0  1 ∪ 5 be the set of dence Φ : 0 → 3(0), where 3(0) is the power set of
patients in the waiting area and (  0 ∪ # be the set 0 and Φ(i) is the set of patients to be seen by the same
of patients in the ED, excluding those sent to tests or physician before patient i. Then,
treatments. ϕ(k)  ϕ(i) for k ∈ Φ(i) and i ∈ 0. (2)
 He, Sim, and Zhang: Data-Driven Patient Scheduling in Emergency Departments
6 Management Science, Articles in Advance, pp. 1–18, © 2019 INFORMS

For patients to be seen by the same physician, the physicians are available to start a consultation. To decide
associated Φ(i) forms a collection of nested sets: for the next patient to be seen at time w(k), we may exploit
i, k ∈ 0, such that ϕ(i)  ϕ(k), we must have information on patients and physicians available before
w(k), including the consultation history of each physi-
Φ(i) ⊂ Φ(k) or Φ(k) ⊂ Φ(i). (3) cian, the identities of waiting patients, the amounts of
time that the physicians have been with their current
The pair of assignment and sequencing decisions (ϕ, Φ) patients, etc. Using the cumulative information, we may
is said to be an admissible schedule if it satisfies (1)–(3). define nonanticipative arrangements through a recursive
We use ! to denote the set of all admissible schedules. procedure. Because the construction of nonanticipative
For a given schedule (ϕ, Φ) ∈ ! and a given realization arrangements is generally complicated, we leave the
of consultation times s  (sij )i∈(,j∈) ∈ 6, the waiting time details to the e-companion. Let 91 be the set of all
of patient k ∈ 0 can be obtained by nonanticipative arrangements. We may obtain an opti-


mal nonanticipative arrangement by solving the problem
wk (s, (ϕ, Φ))  sℓϕ(k) + sℓϕ(k) , (4)
ℓ∈#ϕ(k) ℓ∈Φ(k) max P(wi (s̃, π(s̃)) ≤ τi : i ∈ 0) (6)
where the first sum on the right side is the remaining s.t. π ∈ 91 .
consultation time of the patient being seen by the The P model problem (6) turns out to be intractable
physician (which is zero if #ϕ(k)  Ø) and the second owing in part to the curse of dimensionality induced by
sum is the total consultation time of waiting patients the recursive structure of nonanticipative arrange-
before patient k. ments. To simplify the computation, we may further
We assume that each patient i ∈ 0 has a delay target confine feasible solutions to (5) within the set of static
τi . For a new patient i ∈ 1, τi is the amount of time from arrangements, which is given by
t until his waiting time exceeds the safety limit for his
door-to-provider time. For a returning patient i ∈ 5, τi 90  {π ∈ 9 : π(s1 )  π(s2 ) for s1 , s2 ∈ 6}.
is specified by the scheduling system for his length of
The next proposition states that static arrangements are
stay to meet the mandatory target. We will discuss how
nonanticipative.
to determine delay targets for returning patients in
Section 6. The scheduling system needs to find an Proposition 1. Let 9, 90 , and 91 be the sets of all ar-
admissible schedule for existing patients, under which rangements, static arrangements, and nonanticipative ar-
their waiting times should not exceed the delay targets. rangements, respectively. Then, 90 ⊂ 91 ⊂ 9.
However, because consultation times are random, we Because static arrangements are invariant for all
may not be able to achieve this with complete certainty. realizations of consultation times, finding an optimal
Instead, we seek to maximize the joint probability that static arrangement would be much simpler than
all patient waiting times are within the delay targets. solving (6). It is equivalent to obtaining an optimal
An arrangement is a function π : 6 → !, which maps admissible schedule by solving the following static P
a realization of consultation times to an admissible model problem:
schedule. We use 9 to denote the set of all arrangements.
Under a given arrangement, we may evaluate the joint
max P(wi (s̃, µ) ≤ τi : i ∈ 0) (7)
probability that all waiting times are within the targets
s.t. µ ∈ !.
by (4). Therefore, an optimal arrangement can be ob-
tained by solving the following P model problem The optimal solution µ†  (ϕ† , Φ† ) to (7) specifies the
assignment and sequencing decisions for all waiting
max P(wi (s̃, π(s̃)) ≤ τi : i ∈ 0) (5) patients. In particular, if there is any i ∈ 0, such that
s.t. π ∈ 9. both Φ† (i)  Ø and #ϕ† (i)  Ø hold, patient i will be the
next patient to be seen by physician ϕ† (i) and should be
This formulation, however, cannot be implemented, sent to the physician immediately. This procedure is
because to determine the admissible schedule, one is repeated when a physician completes a consultation or
required to know the realization of s̃ in advance. To fix a new patient arrives in the waiting area finding at least
this issue, we should confine feasible solutions to (5) one free physician. Each time, the scheduling system
within the set of nonanticipative arrangements, which do determines the next patient to be seen for the available
not rely on future information to determine the patients physician.
to be seen when physicians become available. To specify Finding an optimal admissible schedule is still a
a nonanticipative arrangement, we need to determine considerable challenge, although solving (7) is much
the assignment and sequencing decisions in a sequen- simpler than solving (6). Under an admissible sched-
tial manner. Let w(1) ≤ w(2) ≤ ⋯ be the times when ule, evaluating the joint probability in (7) involves
He, Sim, and Zhang: Data-Driven Patient Scheduling in Emergency Departments
Management Science, Articles in Advance, pp. 1–18, © 2019 INFORMS 7

multidimensional integration of many variables, which measure of a hyperrectangle, the patient scheduling
is computationally prohibitive. Nemirovski and Shapiro problem would be more computationally amiable
(2006) pointed out that computing the distribution of under this formulation.
the sum of independent random variables is an NP- We would modify the P model problem (6) to obtain
hard problem. As a result, even finding the distribution a “near-optimal” nonanticipative arrangement. To this
of a patient’s waiting time would be computationally end, let us consider a collection of hyperrectangular
difficult. When the number of waiting patients is large, subsets of 6 given by
we would be unable to obtain the optimal admissible  
schedule for (7) within a reasonable time that is re- *  6 ∩ ∏ [0, dij ] : (dij )i∈(,i∈) ∈ 6 .
i∈(,j∈)
quired for dynamic patient scheduling. Therefore, we
would like to focus on a computationally amiable ap- Our objective is to maximize the joint probability that
proach to obtaining a near-optimal solution. the consultation times are within a hyperrectangular
4. The Hybrid Robust- set 4 ∈ * without exceeding the delay targets:
Stochastic Approach max P(s̃ ∈ 4)
Consider the function wk given by (4) and extend its s.t. wk (s, π(s)) ≤ τk , k ∈ 0, s ∈ 4 (9)
domain to R+|(| |)| × !. Under µ ∈ !, the set 4 ∈ *, π ∈ 91 .
-(µ)  {x ∈ R+|(| |)| : wk (x, µ) ≤ τk for all k ∈ 0} From a robust optimization perspective, * can be
regarded as a family of uncertainty sets for s̃, and the
is a convex polyhedron in |(| · |)| dimensions. Then, we
specific uncertainty set 4 can be adjusted within *
may rewrite (7) as
using different arrangements. The objective function in
max P(s̃ ∈ -(µ)) (9) involves random variables, whereas the constraints
(8)
s.t. µ ∈ !, are based on a robust optimization formulation with an
uncertainty set that is adjustable. Hence, we refer to this
the optimal solution to which is the admissible
formulation as a hybrid robust-stochastic approach.
schedule that maximizes the joint probability of all
Because the representation of nonanticipative ar-
consultation times being within the associated convex
rangements is generally complicated, one may wonder if
polyhedron. Because it is difficult to evaluate P(s̃ ∈ ])
the hybrid formulation is more computationally trac-
for a general polyhedron ] in high dimensions, it
table. To address this concern, let us consider a general
would be computationally excruciating to find the
version of formulation (9), which is not confined to
optimal solution to (8). However, if ] happens to be
scheduling problems for queueing networks. Let σ̃ 
hyperrectangular (e.g., ]  ∏i∈(,j∈) [0, dij ] for some
(σ̃1 , . . . , σ̃M ) be an M-dimensional vector of mutually
dij ≥ 0), the above joint probability can be computed by
independent random variables. For k  1, . . . , M, we use
  Gk to denote the cumulative distribution function of σ̃k .
P s̃ ∈ ]  ∏ P(0 ≤ s̃ij ≤ dij )  ∏ Fij (dij ),
i∈(,j∈) i∈(,j∈)
We use ^ to denote the collection of closed hyper-
rectangular subsets of RM that are bounded from above
because the entries of s̃ are mutually independent. In (i.e., ^  {∏M k1 (−∞, bk ] : (b1 , . . ., bM ) ∈ R }). Let $ be
M

this case, evaluating the joint probability does not in- a nonempty set and & be the collection of functions from
volve the tedious high-dimensional integration. RM to $. Let θ be a function from the product space
The above observation motivates us to consider an RM × $ to RL , such that, for σ  (σ1 , . . . , σM ) ∈ RM and
alternative formulation. Note that, by (8), we intend to ν ∈ $, θ(σ, ν) is nondecreasing in each σk . Then, for
find the admissible schedule with the associated convex a given η ∈ RL , let us consider the following problem:
polyhedron that has the largest probability measure  
induced by s̃. If the probability measure of a convex max P σ̃ ∈ @
s.t. θ(σ, ψ(σ)) ≤ η, σ ∈ @ (10)
polyhedron is large, we may expect the polyhedron to
contain a hyperrectangular subset with a probability @ ∈ ^, ψ ∈ &.
measure that is also large. Conversely, if we can find an Theorem 1 provides a simplified form of (10).
admissible schedule with an associated convex poly-
hedron that contains a “large” hyperrectangle, we may Theorem 1. Let (b⋆ , ν⋆ ) ∈ RM × $ be an optimal solution to
also expect the polyhedron itself to be large. Hence, the following problem:
instead of searching for the admissible schedule that has

M
the “largest” convex polyhedron, we would find the max ln Gk (bk )
admissible schedule with an associated convex poly- k1 (11)
hedron that has the largest hyperrectangular subset. s.t. θ(b, ν) ≤ η, b  (b1 , . . ., bM )
Because it is far easier to evaluate the probability b ∈ RM , ν ∈ $.
 He, Sim, and Zhang: Data-Driven Patient Scheduling in Emergency Departments
8 Management Science, Articles in Advance, pp. 1–18, © 2019 INFORMS

Let @⋆ be the hyperrectangular subset in ^ with boundary representation of nonanticipative arrangements is com-
value b⋆ and ψ⋆ be the constant function with ψ⋆ (σ)  ν⋆ for plicated, finding such an adjustable solution is generally
σ ∈ RM . Then, (@⋆ , ψ⋆ ) is an optimal solution to (10). difficult (Section EC.1 in the e-companion).
This theorem allows us to obtain a simplified form of When the ED is crowded, it may happen that, under
(9), for which a static arrangement is optimal. any admissible schedule, there is at least one patient
whose waiting time will exceed the delay target. In this
Corollary 1. Let (d⋆ , µ⋆ ) ∈ 6 × ! be an optimal solution to case, the hybrid formulation (9) does not have a feasible
the following problem: solution. Because the waiting time given by (4) is in-


creasing with each consultation time, we may deter-
max ln Fij (dij )
i∈( j∈)
mine the feasibility of (9) by examining admissible
(12) schedules when all consultation times take their
s.t. wk (d, µ) ≤ τk , k ∈ 0, d  (dij )i∈(,j∈)
d ∈ 6, µ ∈ !. smallest possible values (i.e., s  (sij )i∈(,j∈)). More spe-
cifically, we may solve the following optimization
Let 4⋆ be the hyperrectangular uncertainty set in * with problem:
boundary value d⋆ and π⋆ be the static arrangement with min α
π⋆ (s)  µ⋆ for s ∈ 6. Then, (4⋆ , π⋆ ) is an optimal solution s.t. wk (s, µ) ≤ τk + α, k ∈ 0, s  (sij )i∈(,j∈) (13)
to (9).
µ ∈ !,
Thanks to the hyperrectangular uncertainty sets, the
hybrid optimization problem (9) has a computation- and we determine the feasibility of (9) by the propo-
ally amiable form given by (12). With these hyper- sition below.
rectangular sets, computing the joint probability in (9)
Proposition 2. Let α⋆ be the minimum value of α given by
is reduced to the double summation in (12) without
(13). Then, the hybrid optimization problem (9) has a feasible
the need for high-dimensional integration. For a given
solution if and only if α⋆ ≤ 0.
uncertainty set, because the worst case occurs only
when all consultation times take their largest values, we
5. A Mixed Integer Program
would be required to examine admissible schedules
The equivalent form (12) can be translated into a mixed
under this single scenario only. It is also worth men-
integer program, which allows us to solve the patient
tioning that existing robust optimization formulations
scheduling problem using existing algorithms.
with adjustable uncertainty sets, such as the budget
For i ∈ (, j ∈ ), and ℓ  1, . . . , |(|, let xijℓ be the binary
of uncertainty by Bertsimas and Sim (2004) and ellip-
variable that indicates the assigned physician and
soidal uncertainty sets by Ben-Tal et al. (2004), may
consultation order of patient i:
not lead to more tractable forms as our hybrid ap-
proach does. 
1 if patient i is the
A feasible solution (4, π) to (9) may result in different xijℓ  ℓth patient to be seen by physician j,
admissible schedules for different realizations of con- 0 otherwise.
sultation times. In this sense, it is an adjustable robust
formulation analogous to the formulation proposed by We reserve |(| positions for each physician so that all
Ben-Tal et al. (2004) for uncertain linear programs. In patients can be accommodated by any physician freely.
general, an adjustable robust formulation is less con- These binary variables must jointly satisfy the fol-
servative than the nonadjustable counterpart, yielding lowing constraints. Because the first position of each
better objective values. Corollary 1, however, implies queue is for the patient who is being seen or to be seen
that we may solve a nonadjustable formulation to immediately by the physician, we have
obtain an optimal solution to (9), which is an admissible xij1  1 for i ∈ #j and j ∈ ). (14)
schedule invariant for all realizations of consultation
times. This is because the worst cases are identical in Furthermore, because returning patients must be seen
both formulations. The hybrid optimization problem by their initial physicians, we have
(9) may also have adjustable solutions, which are

|(|
nonstatic, nonanticipative arrangements. As pointed xijℓ  1 for i ∈ 5j and j ∈ ). (15)
ℓ1
out by de Ruiter et al. (2016), an adjustable solution
may outperform the nonadjustable solution in terms of Under an admissible schedule, each waiting patient can
mean objective values, even if they are both optimal be assigned to only one position:
in the worst case. Hence, there could be adjustable


|(|
solutions to (9) that perform better than the static xijℓ  1 for i ∈ 0, (16)
solution to (12) in patient scheduling. Because the j∈) ℓ1
He, Sim, and Zhang: Data-Driven Patient Scheduling in Emergency Departments
Management Science, Articles in Advance, pp. 1–18, © 2019 INFORMS 9

and each position can accommodate at most one form. By introducing a set of variables {uijℓ ≥ 0 : i ∈ (j ,
patient: j ∈ ), ℓ  1, . . . , |(| − 1}, we may write (20) into two

separate inequalities:
xijℓ ≤ 1 for j ∈ ) and ℓ  1, . . ., |(|, (17)
i∈(j m





uijℓ ≤ xij(m+1) · τi + 1 − xij(m+1) · τ̄
where (j  1 ∪ #j ∪ 5j is the set of patients eligible to be ℓ1 i∈(j i∈0j i∈0j
seen by physician j. Beginning from the first position, we for j ∈ ) and m  1, . . . , |(| − 1 (21)
must assign patients to consecutive positions of each
physician, and therefore, empty positions can appear and uijℓ ≥ xijℓ · dij for i ∈ (j , j ∈ ), and ℓ  1, . . ., |(| − 1.
only at the end of the queue. Because the ℓth position by Because xijℓ ∈ {0, 1} and dij ≤ s̄ij , the latter inequality is
physician j has a patient if and only if equality holds in equivalent to
(17), the above constraint is equivalent to uijℓ ≥ dij − (1 − xijℓ ) · s̄ij


for i ∈ (j , j ∈ ), and ℓ  1, . . . , |(| − 1. (22)
xij(ℓ+1) ≤ xijℓ for j ∈ ) and ℓ  1, . . ., |(| − 1.
i∈(j i∈(j
Then, inequalities (21) and (22) specify the delay con-
(18) straints in canonical form.
One can check that (14)–(18) are equivalent to (1)–(3). In We have obtained the constraints for the patient
other words, each admissible schedule can be deter- scheduling problem given by (14)–(18), (21), and (22),
mined by a set of binary variables {xijℓ : i ∈ (, j ∈ ), ℓ  which are equivalent to the constraints in (12). As suggested
1, . . . , |(|} that satisfies (14)–(18). by the following theorem, we may convert the hybrid
Consider the delay constraints in (12). Put optimization problem into a mixed integer program.

Theorem 2. Let gij (n)  ln Fij (sij (n)) for i ∈ (, j ∈ ) and
τ̄  max s̄ij − min s̄ij , (19) n  1, . . . , Nij . The hybrid optimization problem (9) can be
j∈) i∈(j
i∈(j
written as the following mixed integer program:
where s̄ij is the greatest value that s̃ij can take. Then, τ̄ is



Nij
an upper bound of patient waiting times. Given a set of max yij (n) · gij (n)
binary variables satisfying (14)–(18), we may write the i∈( j∈) n1
delay constraints in (12) into s.t. constraints (14)–(18) and (21)

m




Nij

xijℓ · dij ≤ xij(m+1) · τi + 1 − xij(m+1) · τ̄ uijℓ ≥ yij (n) · sij (n) − (1 − xijℓ ) · s̄ij ,
n1
ℓ1 i∈(j i∈0j i∈0j
i ∈ (j , j ∈ ), ℓ  1, . . . , |(| − 1
for j ∈ ) and m  1, . . . , |(| − 1, (20)


Nij
yij (n)  1, i ∈ (, j ∈ )
where 0j  1 ∪ 5j is the set of waiting patients eligible n1
to be seen by physician j. In this inequality, the sum on xijℓ , yij (n) ∈ {0, 1}, i ∈ (, j ∈ ), ℓ  1, . . ., |(|,
the left side is the time that physician j takes to finish n  1, . . ., Nij
the patients in the first m positions or the waiting time uijℓ ≥ 0, i ∈ (j , j ∈ ), ℓ  1, . . ., |(| − 1.
until the physician begins to serve the (m + 1)st posi- (23)
tion. If there is a patient in the (m + 1)st position, we
have i∈0j xij(m+1)  1, and inequality (20) becomes

m

6. Dynamic Scheduling Using the
xijℓ · dij ≤ xij(m+1) · τi . Hybrid Approach
ℓ1 i∈(j i∈0j
Patient arrivals at the ED form a stochastic process. For
the scheduling system to make sequential decisions ac-
Because the sum on the right side is equal to the delay
cordingly, the proposed hybrid approach must be in-
target, this inequality is the delay constraint for the
corporated in a dynamic scheduling framework. We
(m + 1)st patient. If there is no patient in the (m + 1)st
assume that a decision iteration is triggered when either
position, i∈0j xij(m+1)  0, and the first sum on the right
side of (20) becomes zero. Then, inequality (20) turns out a physician completes a consultation or a new patient
to be comes to the waiting area finding at least one free
physician. Each time, the scheduling system will rec-

m

ommend the next patient to be seen for the available
xijℓ · dij ≤ τ̄,
ℓ1 i∈(j physician.
To solve the dynamic scheduling problem by the
which always holds by the definition of τ̄. For com- hybrid approach, we need to determine delay targets for
putational convenience, let us express (20) in canonical waiting patients when a decision iteration is triggered.
 He, Sim, and Zhang: Data-Driven Patient Scheduling in Emergency Departments
10 Management Science, Articles in Advance, pp. 1–18, © 2019 INFORMS

The major concern for new patients is their door-to- inevitable in solving the patient scheduling problem.
provider times, whereas that for returning patients is Delage and Iancu (2015) discussed time consistency
their lengths of stay. Assume that a decision iteration is issues arising from robust multistage decision making.
triggered at time t. For i ∈ 0, let ti , Di , and Ki be patient
i’s arrival time, safety limit for the door-to-provider time, 7. Data-Based Validation Study
and mandatory target for the length of stay, respectively. We conduct a numerical study to evaluate the per-
For a new patient i ∈ 1, we take the delay target as formance of the hybrid robust-stochastic approach.
τi  Di − (t − ti ), which is the time until the patient’s A set of patient flow data provided by an anonymous
door-to-provider time exceeds the safety limit. Assume hospital is used for validation.
that a patient can return to the same physician at most B This hospital adopts a four-level triage system in the
times. To determine delay targets for returning patients, ED. Levels 1 and 2 are assigned to urgent patients, who
we pick B positive numbers (Ti1 , . . . , TiB ) that satisfy have priority over others. In this ED, urgent patients
Di < Ti1 < ⋯ < TiB < Ki for i ∈ 5, where Tim is interpreted are treated separately in a designated area with ded-
as the mandatory limit for the duration from patient i’s icated personnel and facilities. There are more than 70%
arrival in the waiting area until he is seen by the phy- of patients belonging to level 3. Although their con-
sician for the (m + 1)st time. If a returning patient i ∈ 5 is ditions appear stable, these patients require timely
waiting to be seen for the (m + 1)st time, we take the treatment to resolve their acute symptoms. When the
delay target as τi  Tim − (t − ti ). Imposing additional ED is getting crowded, this group of patients will be the
delay constraints enables us to carry out dynamic most likely to suffer from prolonged waiting.1 In fact,
scheduling by sequentially solving (9). With the extra the crowding of level 3 patients has been the most
delay requirements, returning patients’ waiting times serious problem in the ED. To address this issue, we
for individual consultations can also be maintained at focus on the scheduling of level 3 patients in this
a reasonable level, which may further improve patients’ section. More specifically, we present some empirical
safety and satisfaction. The specific values of these ad- findings from the data of level 3 consultation times
ditional mandatory limits will influence the perfor- in Section 7.1. The computational performance of the
mance of dynamic scheduling. It would be desirable if hybrid approach is compared with that of the SAA
delay targets for returning patients can be adjusted in method in Section 7.2. We assess the hybrid approach
each iteration according to the ED’s congestion. Because for dynamic patient scheduling in Section 7.3, where
finding the optimal delay target for each returning the asymptotically optimal scheduling policy proposed
patient is generally difficult, we will use a heuristic by Huang et al. (2015) serves as a benchmark.
approach to determining these parameters in our
implementation; Section 7.3 has more details. 7.1. Consultation Time Categorization and
Given the delay targets for current waiting patients, Physician Heterogeneity
the scheduling system will first determine the feasi- This set of patient flow data includes the records of
bility of the hybrid optimization problem (9) by solving around 120,000 patient visits to the ED, with over
(13) (which may also be converted into a mixed integer 85,000 visits made by level 3 patients. Each record
program according to the procedure in Section 5). If the contains a series of time stamps, such as the start and
feasible set of (9) is nonempty, the scheduling system end times of triage, consultation, and medical tests,
will solve the mixed integer program (23), the optimal which enable us to reconstruct the patient’s entire path
solution to which specifies the next patient to be seen through the ED. Triage notes and final diagnoses can
for the available physician. If problem (9) turns out also be found from these records.
to be infeasible, the admissible schedule obtained by We divide level 3 patients into two categories. The
solving (13) will be used instead in our implementation. first category includes patients with the most common
In this case, the consultation time of each waiting acute illnesses, such as headache, upper respiratory
patient is assumed to take the minimum value. The tract infection, and acute gastritis, whereas the second
obtained admissible schedule is the one that minimizes category includes all other patients. In practice, the
the longest waiting time of all waiting patients. The categorization of patients should be done at the triage
recommendation for the patient to be seen is made stage according to each patient’s symptoms and vital
based on this admissible schedule. signs. Generally speaking, patients in the first category
We would like to point out that this iterative can be easily identified by the triage nurse, and the
scheduling procedure is myopic in nature. Therefore, diagnosis and treatment of these cases are relatively
admissible schedules obtained from consecutive iter- simple. Nowadays, more and more EDs have imple-
ations could be inconsistent (i.e., successive assignment mented a program known as the fast track, in which
decisions may not satisfy Bellman’s principle of opti- patients having minor illnesses and injuries are iden-
mality). Because a dynamic programming formulation tified by the triage nurse and sent to a dedicated area
would be intractable, time inconsistency is practically for medical care (Sanchez et al. 2006). Hence, patient
He, Sim, and Zhang: Data-Driven Patient Scheduling in Emergency Departments
Management Science, Articles in Advance, pp. 1–18, © 2019 INFORMS 11

categorization can be readily implemented through an Figure 3. (Color online) Boxplot of Consultation Times by
ED’s triage process. We assume that each patient’s Five Physicians
category is known by the scheduling system when the
patient arrives in the waiting area.
In the numerical study, we categorize all level 3 cases
in the dataset into the two groups according to each
one’s diagnosis. There are about 40% of level 3 cases
belonging to the first category. We plot the histograms of
consultation times of the two patient categories in
Figure 2. The mean consultation time of the first cate-
gory is 6.33 minutes, and that of the second category
is 6.94 minutes. In numerical experiments, when all
physicians are assumed to have the same work rates,
these two empirical distributions are used in the
scheduling algorithm as the distributions of consultation
times. Because physicians may be heterogeneous, the
empirical consultation time distributions of each cate-
gory may differ for different physicians. In this case, the
empirical distributions should be generated using the
records of consultation times by each physician. variability in consultation times by physician 3 is much
In the ED of this hospital, there are five to six phy- less than that by any other physician. For a patient
sicians working for level 3 patients in each eight-hour scheduling approach to be relevant to practice, the
shift. Although emergency physicians are required to heterogeneity of physicians must be taken into account.
provide treatment for a wide range of illnesses and
injuries, their expertise and work rates differ from one 7.2. Comparison with the SAA Method
another. Among physicians who completed more than The admissible schedule obtained by solving (9) is in
3,000 cases, we selected five physicians and examined general not the optimal solution to the static P model
the records of patients seen by them. No significant problem (7). Although finding the exact optimal so-
differences have been found among the five patient lution to (7) is difficult, it is possible to use approximate
groups. We also examined the physicians’ consultation approaches, such as the SAA method, to obtain near-
times when they were on day and night shifts. The optimal solutions with reduced computational effort.
distribution of a physician’s consultation times does Let us compare the computational performance of the
not seem to change with time greatly. The boxplot of hybrid approach with that of the SAA method.
consultation times by the five physicians is shown For µ ∈ !, let χ̃(µ) be an indicator random variable
in Figure 3. We can see that the distributions of con- given by
sultation times differ a lot across these physicians. For 
 
instance, the median consultation time by physician 2 χ̃(µ)  1 if wk s̃, µ ≤ τk for all k ∈ 0,
is just one-half of that by physician 5, whereas the 0 otherwise.

Then, E(χ̃(µ))  P(wk (s̃, µ) ≤ τk : k ∈ 0). Let {s1 , . . ., sN },

Figure 2. (Color online) Histograms of Consultation Times where sn  (snij )i∈(,j∈) for n  1, . . . , N, be a sample of
of the Two Categories consultation time vectors independently taken from the
distribution of s̃ and {χ1 (µ), . . . , χN (µ)} be the corre-
sponding realizations of χ̃(µ). By the strong law of large
numbers, the probability of all waiting patients meeting
the delay targets under the admissible schedule µ can be
approximated by N n1 χn (µ)/N. Using this fact, we
formulate an approximate problem for (7) by


N
max zn
n1
s.t. wk (sn , µ) ≤ zn · τk + (1 − zn ) · τ̄n ,
k ∈ 0, n  1, . . . , N
µ ∈ !, zn ∈ {0, 1}, n  1, . . . , N,
(24)
 He, Sim, and Zhang: Data-Driven Patient Scheduling in Emergency Departments
12 Management Science, Articles in Advance, pp. 1–18, © 2019 INFORMS

where sample sizes N  20, 40, 100, and 200. If the sample size

is small, the solutions exhibit great variability in per-
τ̄n  max snij − min snij formance across different realizations, and most of
j∈) i∈(j
i∈(j them are not satisfactory. Increasing the sample size
can stabilize and improve the performance of solutions
is an upper bound of patient waiting times. When at the expense of longer computation time. In solving
zn  0, the inequality in (24) always holds for all k ∈ 0 the mixed integer program for the SAA formulation,
and µ ∈ !; when zn  1, the inequality holds for all the computation time tends to increase when the delay
k ∈ 0 and a given µ ∈ ! if and only if χn (µ)  1. Under targets are shorter. When the pair of delay targets is
a given µ ∈ !, the maximum value that N n1 zn can take (τN , τR )  (30, 50) and (30, 45) minutes, it becomes
must be equal to N χ
n1 n (µ). Therefore, when N is large, difficult to obtain the optimal solution to (24) within
the admissible schedule that maximizes the objective hours for N  200. Instead, we test these two cases with
function in (24) should be a near-optimal solution to (7). N  20, 40, 100, and 150. When the sample size is
Following the procedure in Section 5, one can also N  150, it took up to six hours to obtain the optimal
convert the SAA formulation into a mixed integer solution for (τN , τR )  (30, 45) minutes.
program. We also illustrate the performance of the hybrid
We consider a scenario with 6 physicians and 20 robust-stochastic approach, which is denoted by HRS in
patients in the ED. Eight and 12 patients belong to the Figure 4. The solution to the hybrid formulation does not
first and second categories, respectively. There are 3 depend on specific realizations of a random sample, and
new patients, 12 returning patients, and 5 patients therefore, no variability is present in the performance of
being seen by physicians. Each physician has 2 this approach. Although the objective function in (9) is
returning patients waiting to be seen. We assume that all different from that of the static P model problem (7), the
new patients have the same delay target τN and that optimal solution to the hybrid formulation outperforms
all returning patients have the same delay target τR . All most of the solutions from the SAA realizations. In
patients will leave the ED when they complete their Figure 4, the probability of target attainment by an SAA
current consultations. The six physicians have identical solution is comparable with that by the HRS solution
work capabilities, and therefore, the distribution of each only when the sample size is large. Although there are
consultation time depends on the patient’s category several SAA realizations producing better solutions than
only. All consultation times are sampled from the em- the hybrid approach, there is no SAA solution out-
pirical distributions in Figure 2 according to patients’ performing the corresponding HRS solution by more
categories. than 2% in terms of the probability of target attainment.
The computational performance of the SAA method The SAA method requires much longer computation
is determined mainly by the sample size. With a larger time. To achieve comparable performance, it may take
sample, one may obtain a better solution to the static P hours to obtain a solution from an SAA realization,
model problem (7) at the expense of longer computa- whereas it only takes tens of seconds to obtain the HRS
tion time. We evaluate the SAA formulation (24) with solution. When the sample size is large, the SAA method
different delay targets and different sample sizes. With cannot be used for patient scheduling on a real-time
each set of parameters, we take eight realizations of the basis. Hence, we will use the scheduling policy pro-
random sample and then solve (24) using each re- posed by Huang et al. (2015), which is much more
alization. Computation time is the major concern of this computationally efficient, as the benchmark policy for
step. The obtained admissible schedule is then evalu- dynamic patient scheduling.
ated by Monte Carlo simulation, where consultation
times are resampled from the empirical distributions. 7.3. Experiments on Dynamic Patient Scheduling
In the simulation, the patients are seen by the physi- Now let us evaluate the performance of the hybrid ap-
cians according to the obtained admissible schedule. proach in dynamic patient scheduling. We focus on two
The probability of all patients meeting their delay performance measures: the percentage of patients whose
targets is computed as the performance measure door-to-provider times exceed the safety limits and the
through 1,000 independent simulation runs. percentage of patients whose lengths of stay exceed
We depict the performance of the SAA method in the mandatory targets. Because the major concern is the
Figure 4, where SAA(N) denotes the optimal solution to crowding of level 3 patients, we set the safety limit for all
(24) based on a realization of sample size N. Because the patients’ door-to-provider times to be D  30 minutes
computation time of the SAA method seems to increase and set the mandatory target for all patients’ lengths of
exponentially with the sample size, we use a logarithmic stay to be K  200 minutes. In the numerical experiments
scale for computation time in Figure 4. When the pair below, we generate a stream of 5,000 patients arriving at
of delay targets is taken to be (τN , τR )  (40, 50) and the ED according to a Poisson process with a rate of 15.2
(35, 50) minutes, we test the SAA formulation with patients per hour. For the convenience of simulation, we
He, Sim, and Zhang: Data-Driven Patient Scheduling in Emergency Departments
Management Science, Articles in Advance, pp. 1–18, © 2019 INFORMS 13

Figure 4. (Color online) Computational Performance of the Hybrid Approach and the SAA Method with Different Delay
Targets

assume that all medical tests and treatments by a nurse time limits (T1 , T2 , T3 ) for the durations from a patient’s
take no time. Hence, if a patient returns, he will join the arrival until he starts the second, third, and fourth
queue for the same physician immediately after the cur- consultations, respectively. The selection of these limits
rent consultation. There are four physicians in the ED. will influence the performance of the scheduling al-
All patient consultation times are sampled from the gorithm. On the one hand, increasing these limits will
empirical distributions in Figure 2 by each patient’s accommodate more returning patients within delay
category, with 40% and 60% of patients belonging to targets for their current consultations, thus allowing
the first and second categories, respectively. more new patients to meet the safety limit for door-to-
According to the patient flow data from the hospital, provider times. On the other hand, with larger man-
there are around 75% of patients returning to their datory limits, patients having multiple consultations
physicians at least once before leaving the ED, whereas will be more likely to exceed their length of stay target.
there are less than 4% of patients returning more than To deal with these two concerns, we use a simple
three times. In the numerical experiments, we assume heuristic approach to determining mandatory limits for
that a patient may return at most three times and that returning patients. Because most patients will return at
the probabilities of zero to three returns are 0.25, 0.40, least once, the balance of consultations between new
0.25, and 0.10, respectively. The number of returns is and returning patients can be controlled by dynami-
assumed to be independent of a patient’s category. cally adjusting T1 , the mandatory limit for patients
With these parameters, the ED turns out to be heavily returning for the first time. More specifically, when
loaded with traffic intensity of 93.57%. a decision iteration is triggered, we take
To solve the mixed integer program (23) sequentially, |1|
we need to specify mandatory limits for returning pa- T1  TL + (TU − TL ) , (25)
|1| + |5(1) |
tients when a decision iteration is triggered. Because
all patients have identical requirements for door-to- where TL and TU are two given positive numbers with
provider times and lengths of stay, we identify three TL < TU , |1| is the number of new patients waiting in
 He, Sim, and Zhang: Data-Driven Patient Scheduling in Emergency Departments
14 Management Science, Articles in Advance, pp. 1–18, © 2019 INFORMS

the ED, and |5(1) | is the number of patients waiting for New Patients First. When a physician finishes a con-
their second consultations. If |1|  |5(1) |  0, we set sultation, the scheduling system will assign the new
T1  TL . Following (25), when there are more new patient who has the earliest registration time to the
patients than patients returning for the first time, we physician; if no new patients are available, the sched-
will increase the mandatory limit for patients’ second uling system will assign the returning patient who is
consultations, which allows more new patients to be eligible to be seen by the physician and has the earliest
served within their door-to-provider time limit. When registration time.
there are more patients returning for the first time, we
will decrease T1 , allowing more returning patients to be The Benchmark Policy. This policy was proposed by
served within their length of stay target. In practice, we Huang et al. (2015). When a physician finishes a con-
take both TL and TU to be several times longer than the sultation at time t, the scheduling system will first check
safety limit for door-to-provider times, so that most if there are new patients whose waiting times are about
new patients can be seen quickly. After T1 is de- to exceed or have exceeded the door-to-provider time
termined, we specify two positive numbers Δ1 and Δ2 limit (i.e., if there exists i ∈ 1, such that D − (t − ti ) < ,
and set T2  T1 + Δ1 and T3  T2 + Δ2 . For the conve- where  > 0 is a given excess time). The scheduling
nience of implementation, we assume that Δ1 and Δ2 system will give priority to new patients if such a patient
are invariant in all iterations. We require T3 , the is found, and it will give priority to returning patients
mandatory limit for the fourth consultations, to be otherwise. If a new patient is to be served, the sched-
sufficiently lower than the limit for lengths of stay. uling system will send the one who arrived the earliest
In the first numerical example, we assume that the to the available physician. [Huang et al. (2015) used the
four physicians have identical work capabilities, and shortest deadline first policy for new patients, which is
therefore, the distribution of a patient’s consultation reduced to the FCFS policy when all patients have the
time does not depend on specific physicians. We same door-to-provider limit.] If a returning patient is to
evaluate the hybrid approach with Δ1  Δ2  30 min- be served, the scheduling system will select the one with
utes, reporting the performance for different pairs of the earliest registration time among the returning pa-
(TL , TU ) in Table 1. As one may expect, increasing TL tients who have the shortest expected remaining con-
and TU will generally reduce door-to-provider viola- sultation times. In other words, the returning patient
tions but result in more length of stay violations. We who is the closest to completion will be sent to the
compare the performance of the hybrid approach with available physician. [Huang et al. (2015) adopted a
that of the following scheduling policies. generalized cµ rule for returning patients to minimize
the cumulative congestion cost. Their policy is reduced
Global FCFS. When a physician finishes a consulta- to the closest to exit first policy if a patient’s length of
tion, the scheduling system will send the patient who stay is regarded as his congestion cost.] Under the as-
has the earliest registration time among those eligible to sumption that physicians are homogeneous, this sched-
be seen by this physician. Because medical tests and uling policy is proved asymptotically optimal in
treatments are assumed to be instantaneous, a physi- minimizing the mean length of stay with constraints on
cian will be kept working on each patient until all door-to-provider times.
consultations of this patient are completed. In this case, In the above three scheduling policies, we assume
the global FCFS policy is equivalent to the returning that, when a new patient arrives in the waiting area
patients first policy. finding one or more physicians available, the scheduling

Table 1. Performance Comparison of Door-to-Provider Times (V̄) and Lengths of Stay (L̄)
Under Various Scheduling Policies, with Arrival Rate of 15.2 Patients Per Hour and Four
Homogeneous Physicians

Scheduling policy V̄ L̄ % V̄ > 30 % L̄ > 200 % Violations

Global FCFS 37.64 52.17 45.88 0.46 45.88

New patients first 3.87 74.96 0.10 8.50 8.50
Benchmark,   1 20.59 59.64 36.52 4.46 38.28
Benchmark,   3 19.63 59.91 24.34 4.90 25.76
Benchmark,   6 18.39 61.63 13.66 6.48 15.92
Benchmark,   10 16.39 62.08 5.80 6.64 8.35
SAA (40) 18.66 61.27 22.12 4.96 24.08
HRS, TL  90, TU  120 13.31 63.67 12.80 2.28 13.15
HRS, TL  100, TU  130 13.59 63.55 12.10 2.48 13.09
HRS, TL  105, TU  130 13.40 64.31 12.54 2.44 13.26
He, Sim, and Zhang: Data-Driven Patient Scheduling in Emergency Departments
Management Science, Articles in Advance, pp. 1–18, © 2019 INFORMS 15

system will randomly select a free physician, to whom the door-to-provider requirement, when the bench-
the patient will be sent immediately. mark policy takes   1, 3, or 6 minutes. Note that,
when the excess time is large (e.g.,   6 or 10 minutes),
The SAA Method. When a physician completes a con- the percentage of patients meeting the length of stay
sultation or when a new patient arrives in the waiting target is not satisfactory under the benchmark policy.
area finding at least one free physician, the scheduling The hybrid approach achieves a better balance be-
system will solve (24) to determine the next patient to tween door-to-provider times and lengths of stay,
be seen for the available physician. because it can evaluate the influence of entire consul-
We compare several performance measures in Table 1, tation time distribution, not just that of the first mo-
including the mean door-to-provider time (denoted ments. Although this advantage is gained at a higher
by V̄), the mean length of stay (denoted by L̄), the computational cost, solving the scheduling problem is
percentage of door-to-provider times exceeding 30 still practically efficient under the hybrid formulation.
minutes, the percentage of lengths of stay exceeding 200 In Table 1, the mean length of stay is slightly longer
minutes, and the percentage of patients who experience under our approach than under the benchmark policy.
time violations (i.e., either their door-to-provider times This is because, with door-to-provider time constraints,
exceed 30 minutes or their lengths of stay exceed 200 the benchmark policy is asymptotically optimal in
minutes). Giving priority to returning patients, the terms of mean length of stay. Our approach is aimed at
global FCFS policy yields short lengths of stay but at the complying with mandatory targets for lengths of stay,
expense of long door-to-provider times. If the new pa- and the percentage of patients meeting the targets is
tients first policy is used, the resulting door-to-provider usually regarded as a more important performance
times are short, whereas the lengths of stay turn out to be indicator out of safety concerns.
much longer. When the ED is crowded, neither policy The most important advantage of the hybrid approach
can be used for the ED to meet the stringent time is the capability of patient scheduling in the presence of
constraints. Although the percentage of patients who heterogeneous physicians. When physicians have dif-
experience time violations is relatively low under the ferent work rates, their expertise should be taken into
new patient first policy, the lengths of stay of such account in making scheduling decisions. Consider the
patients are extensively long, resulting in severe con- following scenario. A physician is an expert in treating
gestion in the ED. The benchmark policy is capable of patients in category 1 but not familiar with cases in
striking a balance for these performance measures. As category 2. When the physician becomes available, there
we discussed earlier, this policy depends on consulta- is a new patient in category 2 whose door-to-provider
tion time distributions only through their first moments. time is about to exceed the safety limit. At that moment,
We report the performance of this policy when the should we send the category 2 patient to this “slow”
excess time is taken to be   1, 3, 6, or 10 minutes. physician or keep the category 2 patient waiting and
A larger excess time allows physicians to see more new send a category 1 patient so that the physician can be
patients within the door-to-provider time limit while working at a “fast” rate? In this case, a tradeoff must be
bringing on more length of stay violations. As the excess made between preventing an immediate time violation
time increases, the performance of the benchmark policy and reducing future crowding. The scheduling policy is
becomes more and more analogous to that of the new required to evaluate the consequences of both actions.
patients first policy. Huang et al. (2015) recommended Unfortunately, the benchmark policy does not allow for
that  should be one order of magnitude smaller than the heterogeneous physicians. It is no longer asymptotically
safety limit for door-to-provider times. For example, optimal in reducing the mean length of stay when
with D  30 minutes, we may take   3 minutes as physicians have different expertise.
a practical option. The SAA method is tested with In the second example, the four physicians are as-
sample size being 40 for the overall computation time to sumed to be heterogeneous. Two of them are experts in
be comparable with that of the hybrid approach. Al- treating cases in category 1 but are not good at category
though the SAA method requires longer computation 2; the other two physicians are more experienced in
times than the benchmark policy, its performance is category 2 but are not familiar with category 1. In the
merely comparable with that of the benchmark policy simulation, we generate original consultation times
with   3 minutes. Because no clear advantages are using the empirical distributions according to patients’
shown, the SAA method will not be included in sub- categories, whereas the actual consultation time of
sequent numerical experiments. a patient depends on the specific physician. If the
Denoted by HRS in Table 1, the hybrid robust- physician is an expert in the patient’s category, the
stochastic approach outperforms the benchmark pol- actual consultation time will be 80% of the original
icy in terms of mean door-to-provider time and the time; otherwise, the actual consultation time will be
percentage of length of stay violations in all cases. It 120% of the original time. All other simulation settings
also leads to a greater percentage of patients meeting are the same as in the previous example. In addition to
 He, Sim, and Zhang: Data-Driven Patient Scheduling in Emergency Departments
16 Management Science, Articles in Advance, pp. 1–18, © 2019 INFORMS

the scheduling policies mentioned earlier, we also scheduling system will determine the category to be
consider the following scheduling policy that is mod- served by comparing τ(1) − τ(2) with a threshold value,
ified from the benchmark policy.2 which is proportional to the additional time that the
physician needs for serving a patient in category 2.
The Modified Benchmark Policy. The scheduling sys- More specifically, if τ(1) − τ(2) ≤ δ(C2 − C1 ) for some
tem follows the benchmark policy in deciding whether δ > 0, the earliest new patient in category 1 should be
to serve a new patient. That is, when a physician fin- sent to the physician; otherwise, the earliest new pa-
ishes a consultation at time t, the scheduling system tient in category 2 should be seen. Note that we may
will check if there exists i ∈ 1, such that D − (t − ti ) < , have C2 − C1 ≤ 0, which implies that the slow physician
where  > 0 is the given excess time. Priority will be can also serve a patient in category 2 quickly. In this
given to new patients if such a patient is found and case, the physician will always serve the new patient
given to returning patients otherwise. If a returning who has the earliest registration time. This is reason-
patient will be served, the scheduling system follows able, because it is well known that giving priority to
the benchmark policy, selecting the one who is the patients with short consultation times will mitigate
closest to completion. If a new patient will be served, congestion. If the physician is an expert in category 2,
the physician’s expertise should be taken into account. the scheduling system may follow a similar procedure
Assume that the physician is an expert in category 1. If to determine the next patient to be seen using the same
the new patient who has the earliest registration time coefficient δ. This heuristic may also be extended to
also belongs to category 1, the scheduling system will more than two patient categories.
send this patient to the physician directly. If this patient Under the modified benchmark policy, the scheduling
belongs to category 2, the urgency of serving this pa- system will assign a patient to a slow physician only if
tient should be evaluated against the loss of efficiency the case is relatively urgent, thus making better use of
caused by prolonged consultations. For k  1, 2, let Ck each physician’s expertise. Such a simple heuristic may
be the expected total consultation time (including lead to considerable performance improvement when
possible future returns) of a patient in category k physicians are heterogeneous. In Table 2, the global
provided by this physician. Then, C2 − C1 is the ad- FCFS, new patients first, and benchmark policies do not
ditional time if the physician would serve a patient in differentiate physicians in making scheduling decisions.
category 2 instead of category 1. Let τ(k) be the time Under these policies, a physician’s average work rate is
until the earliest new patient in category k exceeds the identical to that in the previous example, and therefore,
door-to-provider limit (with τ(1)  D if there are no the system’s performance does not show much differ-
new patients in category 1 waiting to be seen). Then, ence. We evaluate the modified benchmark policy for
τ(1) − τ(2) , the difference between the two deadlines, can   3, 6, and 10 minutes. Following the same rule in
be used to measure the relative urgency of serving deciding whether to serve a new patient, the modified
category 2 instead of category 1. Because the new policy does not perform better than the benchmark
patient who has the earliest registration time belongs to policy in terms of door-to-provider time. Instead, it has
category 2, we must have τ(1) ≥ τ(2) . Clearly, it would be the advantage of reducing lengths of stay by allowing
more urgent to serve category 2 if τ(1) − τ(2) is large. The physicians to serve patients in a more efficient way. We

Table 2. Performance Comparison of Door-to-Provider Times (V̄) and Lengths of Stay (L̄)
Under Different Scheduling Policies, with Arrival Rate 15.2 Patients Per Hour and Four
Heterogeneous Physicians

Scheduling policy V̄ L̄ % V̄ > 30 % L̄ > 200 % Violations

Global FCFS 38.69 53.19 46.60 0.36 46.60

New patients first 3.88 72.19 0.12 8.14 8.14
Benchmark,   1 20.74 64.62 36.66 6.06 38.78
Benchmark,   3 20.06 63.32 25.84 6.20 28.92
Benchmark,   6 18.65 63.00 14.38 6.24 18.51
Benchmark,   10 16.39 63.12 11.54 7.04 13.75
Modified benchmark,   3, δ  0.5 21.36 50.24 28.62 1.32 28.93
Modified benchmark,   3, δ  1 21.16 49.14 29.46 1.18 29.46
Modified benchmark,   3, δ  2 21.21 48.98 31.92 1.34 31.92
Modified benchmark,   6, δ  1 20.04 49.11 24.18 1.38 24.42
Modified benchmark,   10, δ  1 17.67 49.06 18.92 2.12 18.92
HRS, TL  90, TU  120 11.63 56.68 10.14 1.82 10.76
HRS, TL  100, TU  130 9.99 56.70 6.76 1.30 6.76
HRS, TL  105, TU  130 9.42 54.23 6.68 1.16 6.72
He, Sim, and Zhang: Data-Driven Patient Scheduling in Emergency Departments
Management Science, Articles in Advance, pp. 1–18, © 2019 INFORMS 17

test δ  0.5, 1, and 2 for   3 minutes, and we test δ  1 2013, Maxwell et al. 2014, Chong et al. 2016), and health
for   6 and 10 minutes. With a larger threshold value, examinations (Baron et al. 2017). Similar problems may
the modified policy will send more patients to fast arise from transportation systems, including taxi dis-
physicians, shortening the mean length of stay. However, patching (Seow et al. 2010), electric vehicle charging
with an increased δ, the modified policy may also management (Yilmaz and Krein 2013), and vehicle routing
postpone more consultations of new patients, although with stochastic demands and time windows (Bertsimas
their door-to-provider times are about to exceed the and van Ryzin 1993, Fisher et al. 1997, Laporte et al. 2002,
safety limit. Therefore, increasing δ may neither shorten Jepsen et al. 2008).
the mean door-to-provider time nor reduce door-to- In a recent paper, Jaillet et al. (2016) extended the
provider violations. One may improve the performance proposed hybrid formulation to a class of satisficing
on door-to-provider times by increasing , which how- problems that are typically computationally intractable.
ever, may worsen the performance on lengths of stay. In that paper, a set of sufficient conditions is identified
As in the previous example, the hybrid approach can for a hybrid formulation to be equivalent to the corre-
achieve a more balanced performance in complying sponding P model. Unfortunately, those conditions do
with the mandatory targets. It outperforms the mod- not apply to the patient scheduling problem in this
ified benchmark policy in terms of door-to-provider paper. How to quantify the loss of optimality from the
time while maintaining a comparable level of length of hyperrectangular approximation of uncertainty sets is
stay violations. We believe that this advantage also an open problem for our future research.
stems from the capability of evaluating the influence of Recent advances in distributionally robust optimiza-
entire distributions. Using the hybrid approach, the tion may also enable us to convert P model problems
scheduling system may send patients to fast physicians into tractable forms. As pointed out by Hanasusanto
with more appropriate timing, thus striking a desir- et al. (2015) and Hanasusanto et al. (2017), one may use
able balance among all performance measures. The distributionally robust formulations to mitigate the in-
e-companion has more simulation results, where at tractability of evaluating high-dimensional integrals
a reasonable computational expense, the hybrid robust- (i.e., when the distribution of a random vector belongs
stochastic approach outperforms other scheduling to certain ambiguity sets, one may obtain the worst
policies in a wide range of parameter regimes. case probability of the random vector being in a given
polyhedron by solving a linear or conic program). By
8. Concluding Remarks those techniques, we may also convert the patient
We proposed a data-driven approach to patient sched- scheduling problem into a mixed integer program.
uling in EDs, where mandatory targets are imposed on Compared with the hybrid approach, a distributionally
patients’ door-to-provider times and lengths of stay. The robust formulation would be particularly useful when
main contribution of this paper is a hybrid robust- either distributional information or patient flow data
stochastic formulation of the patient scheduling prob- are limited.
lem, by which we obtain a near-optimal solution to the
corresponding P model problem at a significantly lower Acknowledgments
computational expense. Using this approach and real- The authors would like to thank the associate editor and the
time patient flow data, we developed a dynamic referees for their thoughtful comments and constructive sug-
scheduling algorithm for making recommendations gestions, which led to a significantly improved paper. In
about the next patient to be seen by each available particular, they are grateful to the anonymous referee who
physician. Our hybrid robust-stochastic approach allows proposed the modified benchmark policy. Any opinions,
findings, and conclusions or recommendations expressed in
for practical features and outperforms existing schedul-
this material are those of the authors and do not reflect the
ing policies in numerical experiments. The capability of
views of the Singapore Ministry of Education or the Singapore
scheduling in the presence of heterogeneous physicians, Government.
in particular, is a major advantage of this approach. In the
future, we may include additional cost structures in
Endnotes
the hybrid robust-stochastic formulation to answer more 1
Level 4 is assigned to nonemergency patients and accounts for
questions arising from patient flow management in EDs. a negligible fraction of visits. We do not consider level 4 patients,
The proposed hybrid formulation may provide a com- because there are no delay requirements for this group.
putationally tractable approach to solving optimiza- 2
The modified benchmark policy was proposed by an anonymous
tion problems in stochastic networks with delay or referee.
throughput time constraints. Such problems often arise
from healthcare systems where service requirements References
are time sensitive, including patient transfer from EDs Armony M, Israelit S, Mandelbaum A, Marmor YN, Tseytlin Y, Yom-
to inpatient wards (Mandelbaum et al. 2012, Shi et al. Tov GB (2015) On patient flow in hospitals: A data-based
2016), ambulance deployment (McLay and Mayorga queueing-science perspective. Stochastic Systems 5(1):146–194.
 He, Sim, and Zhang: Data-Driven Patient Scheduling in Emergency Departments
18 Management Science, Articles in Advance, pp. 1–18, © 2019 INFORMS

Bandi C, Bertsimas D (2012) Tractable stochastic analysis in high Jepsen M, Petersen B, Spoorendonk S, Pisinger D (2008) Subset-row
dimensions via robust optimization. Math. Programming B 134(1): inequalities applied to the vehicle-routing problem with time
23–70. windows. Oper. Res. 56(2):497–511.
Bandi C, Bertsimas D, Youssef N (2015) Robust queueing theory. Laporte G, Louveaux FV, van Hamme L (2002) An integer L-shaped
Oper. Res. 63(3):676–700. algorithm for the capacitated vehicle routing problem with
Baron O, Berman O, Krass D, Wang J (2017) Strategic idleness and stochastic demands. Oper. Res. 50(3):415–423.
dynamic scheduling in an open-shop service network: Case Maglaras C, Van Mieghem JA (2005) Queueing systems with lead-
study and analysis. Manufacturing Service Oper. Management time constraints: A fluid-model approach for admission and
19(1):52–71. sequencing control. Eur. J. Oper. Res. 167(1):179–207.
Ben-Tal A, Goryashko A, Guslitzer E, Nemirovski A (2004) Adjust- Mandelbaum A, Momčilović P, Tseytlin Y (2012) On fair routing from
able robust solutions of uncertain linear programs. Math. Pro- emergency departments to hospital wards: QED queues with
gramming A 99(2):351–376. heterogeneous servers. Management Sci. 58(7):1273–1291.
Bertsimas D, Sim M (2004) The price of robustness. Oper. Res. 52(1): Maxwell MS, Ni EC, Tong C, Henderson SG, Topaloglu H, Hunter SR
35–53. (2014) A bound on the performance of an optimal ambulance
Bertsimas D, Gamarnik D, Rikun AA (2011) Performance analysis redeployment policy. Oper. Res. 62(5):1014–1027.
of queueing networks via robust optimization. Oper. Res. 59(2): McHugh M (2013) The consequences of emergency department
455–466. crowding and delays for patients. Hall R, ed. Patient Flow: Re-
Bertsimas DJ, van Ryzin G (1993) Stochastic and dynamic vehicle ducing Delay in Healthcare Delivery, 2nd ed. (Springer, New York),
routing in the Euclidean plane with multiple capacitated 107–127.
vehicles. Oper. Res. 41(1):60–76. McLay LA, Mayorga ME (2013) A dispatching model for server-to-
Charnes A, Cooper WW (1963) Deterministic equivalents for opti- customer systems that balances efficiency and equity. Manufactur-
mizing and satisficing under chance constraints. Oper. Res. 11(1): ing Service Oper. Management 15(2):205–220.
18–39. Nemirovski A, Shapiro A (2006) Convex approximations of chance
Chong KC, Henderson SG, Lewis ME (2016) The vehicle mix decision constrained programs. SIAM J. Optim. 17(4):969–996.
in emergency medical service systems. Manufacturing Service Plambeck E, Kumar S, Harrison JM (2001) A multiclass queue in
Oper. Management 18(3):347–360. heavy traffic with throughput time constraints: Asymptotically
Connelly LG, Bair AE (2004) Discrete event simulation of emergency optimal dynamic controls. Queueing Systems 39(1):23–54.
department activity: A platform for system-level operations Rabin E, Kocher K, McClelland M, Pines J, Hwang U, Rathlev N,
research. Academic Emergency Medicine 11(11):1177–1185. Asplin B, Trueger NS, Weber E (2012) Solutions to emergency
de Bruin AM, Bekker R, van Zanten L, Koole GM (2010) Di- department ‘boarding’ and crowding are underused and may
mensioning hospital wards using the Erlang loss model. Ann. need to be legislated. Health Affairs 31(8):1757–66.
Oper. Res. 178(1):23–43. Saghafian S, Hopp WJ, Van Oyen MP, Desmond JS, Kronick SL (2012)
de Ruiter FJCT, Brekelmans RCM, den Hertog D (2016) The impact Patient streaming as a mechanism for improving responsiveness
of the existence of multiple adjustable robust solutions. Math. in emergency departments. Oper. Res. 60(5):1080–1097.
Programming A 160(1–2):531–545. Saghafian S, Hopp WJ, Van Oyen MP, Desmond JS, Kronick SL (2014)
Delage E, Iancu DA (2015) Robust multi-stage decision making. Complexity-augmented triage: A tool for improving patient
Aleman DM, Thiele AC, eds. The Operations Research Revolution, safety and operational efficiency. Manufacturing Service Oper.
INFORMS TutORials in Operations Research (INFORMS, Management 16(3):329–345.
Catonsville, MD), 20–46. Sanchez M, Smally AJ, Grant RJ, Jacobs LM (2006) Effects of a fast-
Doytchinov B, Lehoczky J, Shreve S (2001) Real-time queues in heavy track area on emergency department performance. J. Emergency
traffic with earliest-deadline-first queue discipline. Ann. Appl. Medicine 31(1):117–120.
Probab. 11(2):332–378. Seow KT, Dang NH, Lee DH (2010) A collaborative multiagent taxi-
Fisher ML, Jörnsten KO, Madsen OBG (1997) Vehicle routing with dispatch system. IEEE Trans. Automation Sci. Engrg. 7(3):607–616.
time windows: Two optimization algorithms. Oper. Res. 45(3): Shi P, Chou MC, Dai JG, Ding D, Sim J (2016) Models and insights for
488–492. hospital inpatient operations: Time-dependent ED boarding
Gilboy N, Tanabe P, Travers D, Rosenan AM (2011) Emergency Se- time. Management Sci. 62(1):1–28.
verity Index (ESI): A Triage Tool for Emergency Department Care, Sinreich D, Marmor Y (2005) Emergency department operations: The
version 4 (AHRQ Publications, Rockville, MD). basis for developing a simulation tool. IIE Trans. 37(3):233–245.
Green LV, Soares J, Giglio JF, Green RA (2006) Using queueing theory Spacey SA, Wiesemann W, Kuhn D, Luk W (2012) Robust software
to increase the effectiveness of emergency department provider partitioning with multiple instantiation. INFORMS J. Comput.
staffing. Academic Emergency Medicine 13(1):61–68. 24(3):500–515.
Hanasusanto GA, Roitch V, Kuhn D, Wiesemann W (2015) A dis- Weber EJ, Mason S, Carter A, Hew RL (2011) Emptying the corridors of
tributionally robust perspective on uncertainty quantification shame: Organizational lessons from England’s 4-hour emergency
and chance constrained programming. Math. Programming B throughput target. Ann. Emergency Medicine 57(2):79–88.e1.
151(1):35–62. Welch SJ, Asplin BR, Stone-Griffith S, Davidson SJ, Augustine J, Schuur
Hanasusanto GA, Roitch V, Kuhn D, Wiesemann W (2017) Am- J (2011) Emergency department operational metrics, measures and
biguous joint chance constraints under mean and dispersion definitions: Results of the second performance measures and
information. Oper. Res. 65(3):751–767. benchmarking summit. Ann. Emergency Medicine 58(1):33–40.
Hoot NR, Aronsky D (2008) Systematic review of emergency de- Yilmaz M, Krein PT (2013) Review of battery charger topologies,
partment crowding: Causes, effects, and solutions. Ann. Emer- charging power levels, and infrastructure for plug-in electric and
gency Medicine 52(2):126–136. hybrid vehicles. IEEE Trans. Power Electronics 28(5):2151–2169.
Huang J, Carmeli B, Mandelbaum A (2015) Control of patient flow in Yom-Tov GB, Mandelbaum A (2014) Erlang-R: A time-varying queue
emergency departments, or multiclass queues with deadlines with reentrant customers, in support of healthcare staffing.
and feedback. Oper. Res. 63(4):892–909. Manufacturing Service Oper. Management 16(2):283–299.
Jaillet P, Jena SD, Ng TS, Sim M (2016) Satisficing awakens: Models Zhang X, Kamgarpour M, Georghiou A, Goulart P, Lygeros J (2017)
to mitigate uncertainty. Working paper, National University of Robust optimal control with adjustable uncertainty sets. Auto-
Singapore, Singapore. matica 75:249–259.