Machine learning, simulation and optimization in healthcare

This article was downloaded by: [74.217.196.22] On: 09 August 2016, At: 08:08Publisher: Institute for Operations Research and the Management Sciences (INFORMS)INFORMS is located in Maryland, USA

Interfaces

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

Transforming Hospital Emergency Department Workflowand Patient CareEva K. Lee, Hany Y. Atallah, Michael D. Wright, Eleanor T. Post, Calvin Thomas IV, Daniel T.Wu, Leon L. Haley Jr.

To cite this article:Eva K. Lee, Hany Y. Atallah, Michael D. Wright, Eleanor T. Post, Calvin Thomas IV, Daniel T. Wu, Leon L. Haley Jr. (2015)Transforming Hospital Emergency Department Workflow and Patient Care. Interfaces 45(1):58-82. http://dx.doi.org/10.1287/inte.2014.0788

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

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

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

Copyright © 2015, 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, managementscience, and analytics.For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org

http://pubsonline.informs.org

http://dx.doi.org/10.1287/inte.2014.0788

http://dx.doi.org/10.1287/inte.2014.0788

http://pubsonline.informs.org/page/terms-and-conditions

http://www.informs.org

Vol. 45, No. 1, January–February 2015, pp. 58–82ISSN 0092-2102 (print) � ISSN 1526-551X (online) http://dx.doi.org/10.1287/inte.2014.0788

© 2015 INFORMS

THE FRANZ EDELMAN AWARDAchievement in Operations Research

Transforming Hospital Emergency DepartmentWorkflow and Patient Care

Eva K. LeeCenter for Operations Research in Medicine and HealthCare, Atlanta, Georgia 30332;

NSF I/UCRC Center for Health Organization Transformation, Industrial and Systems Engineering, Atlanta, Georgia 30332; andGeorgia Institute of Technology, Atlanta, Georgia 30332, [email protected]

Hany Y. AtallahGrady Health System, Atlanta, Georgia; and Department of Emergency Medicine, Emory University School of Medicine,

Atlanta, Georgia 30322

Michael D. WrightGrady Health System, Atlanta, Georgia 30322

Eleanor T. PostRockdale Medical Center, Conyers, Georgia 30012

Calvin Thomas IVHealth Ivy Tech Community College, Indianapolis, Indiana 46208

Daniel T. Wu, Leon L. Haley Jr.Grady Health System, Atlanta, Georgia; and Department of Emergency Medicine, Emory University School of Medicine,

Atlanta, Georgia 30322

When we encounter an unexpected critical health problem, a hospital’s emergency department (ED) becomesour vital medical resource. Improving an ED’s timeliness of care, quality of care, and operational efficiencywhile reducing avoidable readmissions, is fraught with difficulties, which arise from complexity and uncertainty.In this paper, we describe an ED decision support system that couples machine learning, simulation, andoptimization to address these improvement goals. The system allows healthcare administrators to globallyoptimize workflow, taking into account the uncertainties of incoming patient injuries and diseases and theirassociated care, thereby significantly reducing patient length of stay. This is achieved without changing physicallayout, focusing instead on process consolidation, operations tracking, and staffing. First implemented at GradyMemorial Hospital in Atlanta, Georgia, the system helped reduce length of stay at Grady by roughly 33 percent.By repurposing existing resources, the hospital established a clinical decision unit that resulted in a 28 percentreduction in ED readmissions. Insights gained from the implementation also led to an investment in a walk-in center that eliminated more than 32 percent of the nonurgent-care cases from the ED. As a result of theseimprovements, the hospital enhanced its financial standing and achieved its target goal of an average EDlength of stay of close to seven hours. ED and trauma efficiencies improved throughput by over 16 percentand reduced the number of patients who left without being seen by more than 30 percent. The annual revenuerealized plus savings generated are approximately $190 million, a large amount relative to the hospital’s $1.5billion annual economic impact. The underlying model, which we generalized, has been tested and implementedsuccessfully at 10 other EDs and in other hospital units. The system offers significant advantages in that itpermits a comprehensive analysis of the entire patient flow from registration to discharge, enables a decisionmaker to understand the complexities and interdependencies of individual steps in the process sequence, andultimately allows the users to perform system optimization.

Keywords : systems transformation; systems optimization; machine learning; multiple-resource allocation;mixed-integer program; simulation; decision support; emergency department; acuity level; length of stay;readmission; operations efficiency.

58

Dow

nloa

ded

from

info

rms.

org

by [

74.2

17.1

96.2

2] o

n 09

Aug

ust 2

016,

at 0

8:08

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.

Lee et al.: Transforming Hospital Emergency Department Workflow and Patient CareInterfaces 45(1), pp. 58–82, © 2015 INFORMS 59

Over the past two decades, emergency depart-ment (ED) crowding and delays have become

serious issues for hospitals and health systems inthe United States. ED visits have increased by morethan 2 million per year, characterized by patients whowere older and sicker, and thus required more com-plex, time-consuming workups (i.e., complete medi-cal examinations, including medical history, physicalexam, laboratory tests, X-rays, and analysis) and treat-ments, and by nonurgent patients who use the ED inplace of primary care facilities. The National Hospi-tal Ambulatory Medical Care survey (Centers for Dis-ease Control 2010) reported 130 million ED visits in2010. Despite increased demand, 19 hospitals closedin 2011. In 2012, hospitals reported that more than 40percent of ED patient visits were for nonurgent care,contributing to long waiting times, decreased qualityand timeliness of care, and decreased patient satisfac-tion. Numerous reports have questioned the ability ofU.S. EDs to handle this increasing demand for emer-gency services (Richardson and Hwang 2001; LewinGroup 2002; Derlet 2002; Derlet et al. 2001; Derlet andRichards 2000, 2002; Taylor 2001).

Our project showcases the transformation that canhappen when operations research (OR) is appliedto improve a hospital’s ED operations. Workingwith Grady Memorial Hospital (also referred to asGrady Hospital or Grady), our operations researchersdevised a customizable model and decision supportsystem that couples machine learning, simulation,and optimization to help hospitals improve effective-ness in their EDs. Part of the Grady Health System,Grady Hospital is the fifth-largest safety net hospitalin the United States; these hospitals provide a dispro-portionate amount of care to vulnerable populations(United States Department of Health and Human Ser-vices 2014). Grady implemented our decision supportsystem with beneficial results, such as reduced lengthof stay (LOS), patient waiting times, and readmissions(i.e., repeat admissions related to an initial admis-sion), and improved efficiencies and throughput, allwithout investing additional funds or resources. Sub-sequently, 10 other hospitals implemented our systemand also achieved beneficial results.

Grady’s ED, which is a Level I trauma center, oper-ates the country’s largest hospital-based ambulance

service. Its ED receives more than 125,000 patient vis-its per year, more than 20,000 of whom are traumapatients. Grady provides critical services to Georgia’shealth system. It is home to Georgia’s only poisoncontrol center, the area’s first primary stroke cen-ter, and Georgia’s first cancer center for excellence.Its extended trauma facilities include surgical suites,burn units, the LifeFlight and AngelFlight air med-ical transport programs, Angel II neonatal transportunits, and an emergency medical service ambulanceprogram.

Grady serves a large population of uninsured pa-tients and diverse socioeconomic groups. Of morethan 621,000 annual patient visits, only eight percentof these patients are privately insured (versus 50 per-cent nationally). In 2007, Grady “was in desperateneed of more than $200 million to remain solvent.Grady’s financial collapse has serious consequencesnot just for metro Atlanta—its crisis could reverberateacross the state 0 0 0 Experts say its inefficient customerservice and general administration have created thisfinancial crisis of epic proportions” (de Moura 2007).In the midst of this financial crisis, a new man-agement team came on board to rescue the hospi-tal and transform its operations. The new leadershipwas committed to serious ED system transformationand initiated a joint collaboration with our team ofoperations researchers. Through extensive data collec-tion and vigorous OR analytical advances and recom-mendations, Grady adopted the transformative steps,which included addressing readmissions, quality, andefficiency of care before the Affordable Care Act andits associated penalties were put in place.

The ED crisis is being experienced across the nation.In January 2014, the American College of EmergencyPhysicians ranked [nationwide] ED access as D+ toreflect “that hospitals are not getting the necessarysupport in order to provide effective and efficientemergency care” (American College of EmergencyPhysicians 2014). Grady feels a more-than-averageburden; it treats all patients, regardless if they haveinsurance. For each service that it provides, it incurscosts for which it will be reimbursed only a small por-tion. In addition, many critically ill patients (includingreferrals from other EDs) are routed to Grady becauseof the excellent specialty care that it provides.

Dow

nloa

ded

from

info

rms.

org

by [

74.2

17.1

96.2

2] o

n 09

Aug

ust 2

016,

at 0

8:08

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.

Lee et al.: Transforming Hospital Emergency Department Workflow and Patient Care60 Interfaces 45(1), pp. 58–82, © 2015 INFORMS

The novelty of our OR work has five main aspects.To the best of our knowledge, these have not beenincorporated in previous methods or studies:

1. We optimize within the ED system simulation,rather than relying on a scenario-based method, sothat the results more closely approach global opti-mization. The global solution involves aligning andconsolidating operations, optimizing staffing, andoptimizing processes.

2. We dynamically and stochastically incorporatetreatment patterns and patient characteristics withinan agent-based simulation, while focusing on EDoperations and quality improvement.

3. We model ED readmissions using data thatsimultaneously encompass demographics, socioeco-nomic status, clinical information, hospital opera-tions, and disease behavioral patterns.

4. We explicitly model the interdependencies to andfrom the ED with numerous other hospital depart-ments, capturing inefficiencies in those processes.

5. We integrate machine learning within the simu-lation-optimization framework.We also note that in our work, all medical terms andrelated metrics are defined as is customary in themedical community.

From a hospital’s perspective, healthcare leadershave acknowledged that this work advances ED oper-ations in several ways, which we describe in theBenefits and Impacts section. From an OR perspec-tive, the collaboration this project engendered andthe challenges it presented have led to both theoreti-cal and computational advances in optimization andsimulation.

BackgroundCrowded ED conditions have sparked research onseveral fronts. Eitel et al. (2010) discussed differ-ent methods for improving ED quality and flow,including demand management, critical pathways,process mapping, emergency severity-index triage,bedside registration, and Lean and Six Sigma man-agement methods (Bahensky et al. 2005). Popovichet al. (2012) developed a volume-driven protocol andimplemented it through the use of published evi-dence, which focused on essential endpoints of mea-surement. Wiler et al. (2010) evaluated interventions,such as immediate bedding, bedside registration,

advanced triage, physician and (or) practitioner attriage, and dedicated fast-track service lines, all ofwhich are considered potential solutions to streamlinethe front-end processing of ED patients. Ashby et al.(2008) optimized patient flow throughout the inpa-tient units, while modeling and observing the impactson other interdependent parts of the hospital, such asthe ED and operating rooms. Kolker (2008) tried toestablish a quantitative relationship between ED per-formance characteristics, such as percentage of timeon ambulance diversion and the number of patientsin queue in the waiting room, and the upper limitsof patient LOS. Moskop et al. (2009) identified anddescribed operational and financial barriers to resolv-ing the crisis of ED crowding; they also proposeda variety of institutional and public policy strate-gies to overcome those barriers. Nugus et al. (2011)used an ethnographic approach that involves directobservation of on-the-ground behaviors, observinginteractions among physicians and nurses, emergencyclinicians, and clinicians from other hospital depart-ments to identify indicators of and responses topressure in the day-to-day ED work environment.DeFlitch et al. (2007) reported provider-directed queu-ing for improving ED operations. McCarthy et al.(2009) used discrete-time survival analysis to deter-mine the effects of crowding on ED waiting room,treatment, and boarding times (i.e., the time spent inthe ED after the decision has been made to admit thepatient to the hospital) across multiple sites and acu-ity levels. Sturm et al. (2010) identified predictors thatcan influence nonurgent pediatric ED utilization.

Challenges and ObjectivesAlthough some of the ED advances have been suc-cessful, the improvement is often not sustainable, orit redirects inefficiencies from one area of the ED toanother, or to other hospital divisions. Poor resultsfrom these approaches are partly because the requisitedata are very time consuming to collect, often resultingin poor data being entered into a model. In addition, amodel may be flawed if important elements and sys-tem dependencies are overlooked in its design.

Readmissions are a key challenge in ED perfor-mance. In particular, avoidable readmissions (i.e.,readmissions resulting from an adverse event that

Dow

nloa

ded

from

info

rms.

org

by [

74.2

17.1

96.2

2] o

n 09

Aug

ust 2

016,

at 0

8:08

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


occurred during the initial admission or from inappro-priate care coordination following discharge) throughthe ED have become a major burden on the U.S. healthsystem; see Minott (2008). Recent research shows thatnearly one in five patients are readmitted to the dis-charging hospital within 30 days of discharge; thesereadmissions accounted for $17.8 billion in Medicarespending in 2004 (Osei-Anto et al. 2010).

Numerous studies have been conducted to iden-tify frequently readmitted patients’ characteristics andconstruct patient profiles to aid hospitals in predictingthese patients. These studies have identified a numberof demographic and clinical factors that are thought tosignificantly correlate with readmission. Other factorsconcerning hospital operations have also been inves-tigated. Various statistical tools have been used toidentify patient factors that are associated with read-missions (Allaudeen et al. 2011a, Billings et al. 2006,Hasan et al. 2010, Kirby et al. 2010). Westert et al.(2002) conducted an international study, includingthree U.S. states and three countries, to find patternsin the profiles of readmitted patients. The findingsare divided into demographic and social factors, clin-ical factors (Billings et al. 2006, Southern et al. 2004),and hospital operations factors (Benbassat and Taragin2000, Davidson et al. 2007, Joynt et al. 2011, Scuteriet al. 2011, VanSuch et al. 2006, Westert et al. 2002).A study of 26 readmission risk-prediction models con-cluded that after reviewing 7,843 citations, none ofthe models analyzed could suitably predict future hos-pital readmissions (Kansagara et al. 2011). Allaudeenet al. (2011b) noted that healthcare personnel couldnot accurately predict the readmission of patients dis-charged from their own hospitals; however, conclu-sions from these studies may be premature, giventhat much of the analyses were performed via logis-tic regression on only subsets of data. We recentlypublished a readmission study in which, for the firsttime, a predictive model can incorporate comprehen-sive factors related to demographics and socioeco-nomic status, clinical and hospital resources, opera-tions and utilization, and patient complaints and riskfactors for global prediction (Lee et al. 2012a). Ourapproach empowers healthcare providers with goodpredictive capability, which we generalized for thisGrady study.

The Affordable Care Act, its influence on Medicaidand Medicare payments, the high cost of emergencycare, the persistent nonurgent visits, and the penaltiesimposed because of inappropriate readmissions andhospital-related health problems demand transforma-tion of ED patient care and workflow.

This project focuses on large-scale systems model-ing and decision analytics for modeling and optimiz-ing the workflow for an ED. Specifically, we aim toimprove workflow, reduce wait time, improve qual-ity and timeliness of care, and reduce the numberof avoidable readmissions. Although most studiesincorporate simulation to model ED operations andperform scenario-based improvement (e.g., Medeiroset al. 2008), we believe that our model is the firstto intertwine machine learning, simulation, and opti-mization into one system in which (1) the ED patientcharacteristics are analyzed and patterns uncovered,and (2) operations and workflow are modeled andresources optimized within the system to achieve thebest performance outcome.

Grady began an ED process transformation with ourOR team in 2008. At that time, the average patientLOS in its ED exceeded 10 hours. LOS represents thetime between a patient’s arrival at the ED and the timethat patient is discharged from the ED or admitted tothe hospital. Thus, LOS includes the door-to-providertime, the time the patient waits for the service andreceives care, and the boarding time. Hence, LOS isoften dominated by long stretches of nonservice times.Grady’s goal was to achieve a LOS of close to sevenhours and reduce its readmissions rate by 25 percent.We refer to the period from the beginning of the studyin 2008 to the time of the sustained improved perfor-mance (July 2011) as Phase I. As a result of the Phase Iimprovements, the hospital was able to use sponsoredfunds to open a walk-in center for low-acuity patients,further driving down costs and LOS (Williams 2011).In addition, the implementation of an electronic med-ical record (EMR) system in October 2010 has enabledthe administration to better track hospital operations.Because of the alternative care options and the addi-tion of a new dedicated 15-bed trauma center, thedynamics of ED patient visits have changed. PhaseII captures the period of ED advances from 2011 tothe time of this writing. This paper summarizes theOR analytic, system-driven advances in the ED and

Dow

nloa

ded

from

info

rms.

org

by [

74.2

17.1

96.2

2] o

n 09

Aug

ust 2

016,

at 0

8:08

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


their associated performance outcomes during thesetwo phases. By design, the two phases overlap.

Methods and Design of StudyThe study involves seven major steps, as follows:

1. Process mapping of ED patient and service work-flow via structured interviews and objective processobservations.

2. Time-motion studies of patient arrival, serviceprocesses, and analysis of hospital data.

3. Development of a machine-learning predictiveanalytic framework to process data and predict patientcharacteristics, complaint types, and admission andreadmission patterns.

4. Development of a computerized simulation-optimization system.

5. System optimization and comparison of optimalsystem performance with existing operations.

6. Determination of actionable recommendationsfor implementation.

7. Evaluation of system improvements.Figure 1 highlights the study schema and the inter-

dependencies of our methods. The human-centeredcomputational modeling environment comprises dataanalytics served by innovative OR predictive decisiontools. We simultaneously explore patterns of patient

ED: Stakeholders'characteristics and

objectives;arrival rates;service pdfs;workflow;resources;

decision process;

Procedural andprocess

benchmarks

Electronicmedical records

Machine learning,predictiveanalytics

Establish processand systems

interdependencies

Simulation-optimizationdecision framework

(i) Validation(ii) Systems optimization

Hospitalhistorical data

Process maps,time-motion study

Refine, nextphase

Implement anddocument

performance

Prioritize andobtain buy-in

Observationalrecommendations

Big-Data Analytic Framework

OR-Predictive AnalyticDecision Framework

Synthesize actionablerecommendations

OUTSIDE UNITS:IN AND OUT OF ED

Figure 1: (Color online) This figure shows the study schema and interdependencies of the analytic frameworkthat we use. These interdependencies are crucial to achieving a valid description of the actual processes.

behavior and care characteristics, provider decisionand process workflow, facility layout design, andstaffing, where resource allocation, cognitive humanbehavior, and care patterns are optimized globally forbest outcomes, as measured by LOS and readmissions.Uncovering patterns in patient care helps to appro-priately align resources with demands, and enablesproviders to better anticipate needs. Exploring facilitydesign provides decision makers with the envisionedimprovement before they embark on an expensive lay-out redesign effort.

ED Workflow and ServicesPatients who visit the ED for care are evaluated firstat the triage area to determine the severity of theirinjuries and (or) conditions. They are assigned anacuity level based on the emergency severity index(ESI), a five-level index for prioritizing ED patientsfor care, ranging from level 1 (emergent and requir-ing multiple resources) to level 5 (nonurgent and leastresource intensive). The ESI is unique among triagetools, because it categorizes ED patients by both acuityand resource needs.

At Grady, the blue zone is used to treat high-acuitypatients (levels 1 and 2) and all prisoners, exceptthose with significant trauma. (Note that a detentionarea for prisoners is located inside the blue zone and

Dow

nloa

ded

from

info

rms.

org

by [

74.2

17.1

96.2

2] o

n 09

Aug

ust 2

016,

at 0

8:08

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


prisoner patients are registered in the blue-zone treat-ment area). A major resuscitation room anchors thisarea, which also includes eight critical care rooms,seven respiratory isolation rooms, and several general-patient-care areas. The red zone is used to treat generalpatients. This area has general-care rooms, an ortho-pedic room, a gynecology evaluation room, and aneye, ear, nose, and throat room. All rooms in the blueand red zones are capable of cardiac monitoring. In2010, based on our Phase I results, Grady added a clin-ical decision unit. This unit provides an alternative toadmission to the main hospital by providing obser-vation services for those patients who have alreadyreceived treatment in the ED. All patients in the clin-ical decision unit are evaluated by a case managerwho helps coordinate care, provide education, andensure appropriate follow-up. A patient who is notimproving is sent to the hospital’s main building foradmission.

The mission of the patient ambulatory care expressarea (PACe) is to treat patients with relatively minorconditions. The PACe facility operates 24 hours a day,seven days a week, and is staffed primarily by nursepractitioners and physician assistants. The traumacenter is designed to treat patients with trauma lev-els 1, 2, and 3, as categorized by the American TraumaSociety (2014). Trauma operating rooms are staffed24 hours a day year-round.

Space/beds Worker type

Patient type 2008 2010 Attending physicians Mid-level providers Nurses

Triage All Yes Nurse practitioners, Yesphysician assistants

Blue zone High-acuity patients and all prisoners without 34 37 Yes Residents Yes(acuity levels 1 and 2) significant trauma

Red zone General patients 25 21 Yes Residents Yes(acuity levels 2 and 3)

Clinical decision unit Treated ED patients who need observation 0 7 Yes No YesPACe Patients with relatively minor conditions 8 8 No Nurse practitioners, Yes

(acuity levels 4 and 5) physician assistantsTrauma center Patients who meet either trauma level 1, 2, 3 4 15 Yes Residents Yes

criteria in addition to any child involved ina traumatic accident, any patient arriving on abackboard, all gunshot and stab victims, andpatients with complex extremity injuries or burns

Table 1: The table shows zones, patient and worker types, and resource availability.

Patients arriving in an ambulance or other vehi-cle enter the ED through the ambulance arrival area,which is separate from the walk-in area. Here, patientsdetermined to be ESI level 1 or 2 will be triaged andsent directly to the blue or red zone. Level 3 and 4patients are triaged, sent to the ED waiting room, andenter the same queue as the patients in the walk-inarea to wait for a bed in either zone. Walk-in patients,some of whom may not be assigned an acuity level, aretreated by the walk-in triage physician and dischargedfrom there.

Table 1 summarizes the ED patient care and re-sources, excluding the walk-in. Figures 2(a)–2(d) showthe workflow process maps at the start of our study.

Data Collection and Time-Motion StudiesIn this section, we discuss the two phases of our study.

Phase I. From August 2008 through February 2009,multiple trained observers collected ED data byreviewing files and charts and conducting interviewsand time-motion studies related to services at variousstations, as guided by the process maps. The data col-lected in this manner contain 45,983 data fields cov-ering 2,509 patients. In addition, the hospital main-tained vital statistics, including acuity level, LOS, anddischarge data. Grady also provided readmission sta-tus for 42,456 patients. Furthermore, we received morethan 40,000 individual service times for laboratoryturnaround—the amount of time between the time a

Dow

nloa

ded

from

info

rms.

org

by [

74.2

17.1

96.2

2] o

n 09

Aug

ust 2

016,

at 0

8:08

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


BL

UE

BE

DA

ssig

nmen

tB

LU

E M

idA

sses

smen

tB

LU

E L

ab?

BL

UE

Lab

Ord

er 1

BL

UE

Lab

Wai

t 1B

LU

E L

abO

rder

2B

LU

E L

abW

ait 2

BL

UE

Doc

ter

Eva

luat

ion

2

BL

UE

Mor

eL

ab?

CD

U o

r no

tC

DU

CD

U A

dmit

orD

isch

arge

BL

UE

Adm

it or

Dis

char

geB

LU

E R

ecov

ery

BL

UE

Dis

char

geB

LU

E B

edC

lean

sing

CD

U D

isch

arge

CD

U A

dmitt

ance

BL

UE

Adm

ittan

ceB

LU

E A

dmitt

ance

Dis

posa

l

CD

U B

edC

lean

sing

CD

U D

ispo

sal

BL

UE

Dis

posa

l

BL

UE

Doc

tor

Eva

luat

ion

1 Figu

re2(

a):

The

flow

char

tsh

ows

the

wor

kflow

proc

ess

map

for

the

blue

zone

.

RE

D B

ED

Ass

ignm

ent

RE

D M

idA

sses

smen

tR

ED

Lab

?

RE

D L

ab O

rder

1R

ED

Lab

Wai

t 1R

ED

Lab

Ord

er 2

RE

D L

ab W

ait 2

RE

D D

octe

rE

valu

atio

n 1

RE

D M

ore

Lab

?R

ED

Adm

it or

Dis

char

ge

RE

D D

octe

rE

valu

atio

n 2

RE

D R

ecov

ery

RE

D A

dmitt

ance

RE

D A

dmitt

ance

Dis

posa

l

RE

D D

isch

arge

RE

D B

edC

lean

sing

RE

D D

ispo

sal

Figu

re2(

b):

The

flow

char

tsh

ows

the

wor

kflow

proc

ess

map

for

the

red

zone

.

PAC

BE

DA

ssig

nmen

tPA

C M

idA

sses

smen

tPA

C L

ab?

PAC

Lab

Ord

er 1

PAC

Lab

Wai

t 1

PAC

Doc

tor

Eva

luat

ion

1PA

C D

isch

arge

PAC

Bed

Cle

ansi

ngPA

C D

ispo

sal

Figu

re2(

c):

The

flow

char

tsh

ows

the

wor

kflow

proc

ess

map

for

the

PACe

.

MT

C B

ED

Ass

ignm

ent

MT

C M

idA

sses

smen

tM

TC

Lab

?M

TC

Doc

tor

Eva

luat

ion

MT

C M

ore

Lab

?M

TC

Adm

it or

Dis

char

ge MT

C D

octo

rE

valu

atio

n 2

MT

C L

ab W

ait 2

MT

C L

ab O

rder

2M

TC

Lab

Wai

t 1M

TC

Lab

Ord

er 1

MT

C R

ecov

ery

MT

C D

isch

arge

MT

C B

edC

lean

sing

MT

C A

dmitt

ance

Dis

posa

lM

TC

Adm

ittan

ce

MT

C D

ispo

sal

Figu

re2(

d):

The

flow

char

tsh

ows

the

wor

kflow

proc

ess

map

for

the

traum

ace

nter

.

Dow

nloa

ded

from

info

rms.

org

by [

74.2

17.1

96.2

2] o

n 09

Aug

ust 2

016,

at 0

8:08

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


laboratory receives a specimen and the time the resultsare available.

Phase II. In Phase II, data from 16,217 patient vis-its from October 28, 2010 to December 31, 2010were pulled from the EMR system. For each visit,data include patient information, ED admission time,hospital discharge time, acuity level, ED zones, diag-nosis, and insurance type. The EMRs also includetime stamps for relevant events, including registra-tion, triage, laboratory orders and results, doctorassessment, observation, and discharge. We supple-mented the EMR ED data with observations and time-motion studies at the triage and registration areas, andsampled treatment and wait times inside the rooms byshadowing various care providers.

By identifying the responsibilities of each type ofworker via shadowing in the ED and reviewing theEMR system patient timeline, we found variabilityand randomness in arrival and treatment processesand workers’ responsibilities. The variability emergesfrom the delivery practices of the nurses, mid-levelpractitioners, and attending physicians. Our modelmay not describe each case in the ED; however, it rep-resents more than 93 percent of the cases.

Machine Learning for Predicting PatientCharacteristics and Return PatternsArmed with comprehensive data, we first devel-oped machine-learning techniques to uncover patientcharacteristics, including resource needs, treatmentoutcome, LOS, and readmission patterns, and to estab-lish predictive rules. A significant contribution of ourwork is that it is the first study in which demo-graphics, socioeconomic status, clinical information,hospital operations, and disease behavioral patternsare employed simultaneously as attributes within amachine-learning framework.

The computational design of our machine-learningframework utilizes a wrapper approach; specifically,we apply pattern recognition based on our recentadvances on text mining for unstructured clinicalnotes to the input attributes (Hagen et al. 2013).Next, we couple a combinatorial attribute-selectionalgorithm with a discriminate analysis via a mixed-integer program (DAMIP) learning and classifica-tion module. The attribute selection, classification,and cross-validation procedures are wrapped so that

the attribute-selection algorithm searches through thespace of attribute subsets using the cross-validationaccuracy from the classification module as a measureof goodness. The small subset of attributes returnedfrom the machine-learning analysis can be viewed ascritical patient and clinical and (or) hospital variablesthat drive service characteristics. This provides feed-back to clinical decision makers for prioritization andintervention of patients and tasks.

In the ED study, entities correspond to patients.The input attributes for each patient include com-prehensive demographics, socioeconomic status, clin-ical information, hospital resources and utilization,and disease behavioral patterns. The machine learninguncovers patient disease patterns, associated resourceneeds, and factors influencing treatment characteris-tics and outcome. For readmission, there are two sta-tuses for patients: they come back to the hospitalfor visits (return group), or they do not come back(nonreturn group). The classification aims to uncoverfrom the set of all attributes a set of discriminatoryattributes that can classify each patient into the returnor nonreturn group. We seek to identify the rule thatoffers the best predictive capability.

In this supervised classification approach, the sta-tus of each patient in the training set is known. Thetraining set consists of a group of patients extractedfrom the hospital database whose status (e.g., returnedwithin 72 hours after the first visit or within 30 daysafter the first visit) is known. The training data areinput into the machine-learning framework. Throughthe attribute-selection algorithm, a subset of attributesis selected to form a classification rule. This rule is thenused to perform 10-fold cross-validation on the train-ing set to obtain an unbiased estimate.

In 10-fold cross-validation, the training set is ran-domly partitioned into 10 roughly equal subsets. Ofthe 10 subsets, one subset is retained as the valida-tion data for testing the model, and the remaining ninesubsets are used as training data. The cross-validationprocess is then repeated 10 times (the folds), with eachof the 10 subsets used exactly once as the validationdata. The 10 results from the folds can then be aver-aged to produce an unbiased estimation. The advan-tage of this method over repeated random subsam-pling is that all observations are used for both training

Dow

nloa

ded

from

info

rms.

org

by [

74.2

17.1

96.2

2] o

n 09

Aug

ust 2

016,

at 0

8:08

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


and validation, and each observation is used exactlyonce for validation.

To gauge the predictive power of the rule, weperform blind prediction on an independent set ofpatients; these patients have never been used inattribute selection. We run each patient through therule, which returns a status. We then give the hospitalpersonnel the predicted status of each patient, whichthey check against the patient’s actual status. Hence,we always compare our prediction with the actual out-come in measuring predictive accuracy.

Once our machine-learning system sends a triggerthat a particular patient is highly likely to return, anexpert human (usually a nurse) places this patienton a to-observe list. That our predictions are not 100percent accurate is understandable; however, the firstpass is critical because it narrows down the return toa very small subset of patients, allowing the humanexpert to focus on them to determine which patientsin this selected set should be observed in the clinicaldecision unit. Learning is continuous because humanexperts may identify attributes that they will use intheir second pass of selection. These attributes willthen be incorporated into our system for learning andrefinement. Lee et al. (2003), Lee (2007), Lee and Wu(2007), Brooks and Lee (2010), and (2014) detail theDAMIP modeling and its theoretical and computa-tional contributions. We include a mathematical for-mulation in the appendix. Briefly, DAMIP employs a0-1 variable to denote if an entity is classified correctly.The model includes features that do not simultane-ously exist in other classification models: (1) a mathe-matical expression that transforms the high-dimensionattribute space into the group space to describe thegroup to which an entity will be classified; (2) a re-served-judgment region that handles entities whosegroup status is difficult to determine correctly to avoidovertraining and facilitate multistage classification;and (3) constraints on the percentage of misclassifica-tions in each group. The model seeks to maximize thenumber of correct classifications.

DAMIP’s special characteristics include the follow-ing: (1) the resulting classification rule is stronglyuniversally consistent, given that the Bayes opti-mal rule for classification is known (Brooks and Lee2010); (2) the misclassification rates using the DAMIP

method are consistently lower than with other clas-sification approaches when tested on both simulatedand real-world data; (3) the DAMIP classificationrules appear to be insensitive to the specification ofprior probabilities, yet capable of reducing misclas-sification rates when the number of training enti-ties from each group is different; and (4) the DAMIPmodel generates stable and robust classification rulesregardless of the proportions of training entities fromeach group (Brooks and Lee 2010, 2014; Lee 2007;Lee and Wu 2007; Lee et al. 2003; Hagen et al. 2013;Koczor et al. 2013).

Note that, mathematically, DAMIP is proven tobe NP-complete (Brooks and Lee 2010, 2014). Wesolved the instances for this ED study using advancesin hypergraphic theory (Lee and Maheshwary 2013,Lee et al. 2014). Based on previous studies, theDAMIP approach generally works better than othermachine-learning methods for unbalanced and het-erogeneous data, as we encountered in this project(Lee et al. 2012a).

The Computerized ED System Workflow ModelTo establish a framework for modeling and optimizingthe ED workflow, including ED processes and depen-dencies on other hospital divisions when dischargingpatients from the ED, we use the RealOpt© simulation-optimization decision support environment (Lee et al.2006a, b, 2009, 2011, 2012b, 2013a; RealOpt 2012).RealOpt was developed at Georgia Tech for thepurpose of optimizing operations, throughput, andresource allocation for public health resources, in par-ticular for emergency response and public health med-ical preparedness. It includes easy-to-use drawingtools to permit users to enter the workflow via mouseclicks and keystrokes. It also allows incorporation ofthe stochastic nature of human behavior (both theservers and patients) in workflows and processes, andprovides a method to model fatigue and stress fac-tors. In the background, it translates the workflow intoa computerized simulation model in which resourcescan be optimized to achieve the best throughput andsystem performance. Figure 3(a) shows the (simpli-fied) clinic workflow and service zones for Grady’sED, as entered into RealOpt via its graph-drawingpanel.

Figure 3(b) shows the average total time to admit apatient from the ED to different units of the hospital.

Dow

nloa

ded

from

info

rms.

org

by [

74.2

17.1

96.2

2] o

n 09

Aug

ust 2

016,

at 0

8:08

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Var

iabl

e

Arr

ival

Acu

ity

Acu

ityL

evel

_1

Acu

ityL

evel

_2 Acu

ityL

evel

_3

Acu

ityL

evel

_4

Acu

ityL

evel

_5

AC

_1_A

rriv

al

AC

_2_A

rriv

al

AC

_3_A

rriv

al

AC

_4_A

rriv

al

AC

_5_A

rriv

al

AC

1_If

Tra

uma

AC

2_If

Tra

uma

AC

3_If

Tra

uma

AC

4_If

Tra

uma

AC

5_If

Tra

uma

MT

C B

ed D

elay

MT

C B

edA

ssig

nmen

tM

TC

Mid

Ass

essm

ent

MT

C L

ab?

MT

C D

octo

rE

valu

atio

nM

TC

Mor

eL

ab?

MT

C A

dmit

orD

isch

arge

MT

C R

ecov

ery

MT

C D

isch

arge

MT

C B

edC

lean

sing

MT

C D

ispo

sal

MT

C A

dmita

nce

Dis

posa

lM

TC

Adm

itanc

eM

TC

Doc

tor

Eva

luat

ion

2M

TC

Lab

Wai

t 2M

TC

Lab

Ord

er 2

MT

C L

ab W

ait 1

MT

C L

ab O

rder

Zon

e or

Wal

kin

Zon

e_W

alki

n

Wal

kin

Reg

istr

atio

n

Zon

e_A

m

Am

bula

nce

Reg

is/T

riag

e

PAC

E B

edD

elay

RE

D B

ed D

elay

PAC

E B

edA

ssig

nmen

t

RE

D B

edA

ssig

nmen

t

PAC

E M

idA

sses

smen

t

PAC

E L

abO

rder

1

RE

D L

abO

rder

1

BL

UE

Lab

Ord

er 1

BL

UE

Lab

Wai

t 1B

LU

E L

abO

rder

2B

LU

E L

abW

ait 2

BL

UE

Doc

tor

Eva

luat

ion

2B

LU

E A

dmit

orD

isch

arge

BL

UE

Rec

over

y

BL

UE

Adm

ittan

ceB

LU

E A

dmitt

ance

Dis

posa

l

BL

UE

Dis

char

ge

RE

D L

abW

ait 1

RE

D L

abO

rder

2R

ED

Lab

Wai

t 2R

ED

Doc

tor

Eva

luat

ion

2R

ED

Adm

ittan

ceR

ED

Adm

ittan

ceD

ispo

sal

PAC

E L

abW

ait 1

RE

D B

edA

sses

smen

t

PAC

E L

ab?

RE

D L

ab?

PAC

E D

octo

rE

valu

atio

n 1

RE

D D

octo

rE

valu

atio

n 1

RE

D M

ore

Lab

?R

ED

Adm

it or

Dis

char

geR

ED

Rec

over

yR

ED

Dis

char

geR

ED

Bed

Cle

ansi

ngR

ED

Dis

posa

l

CD

U D

ispo

sal

BL

UE

Dis

posa

lB

LU

E B

edC

lean

sing

CD

U B

edC

lean

sing

CD

U D

isch

arge

CD

U A

dmitt

ance

CD

U A

dmit

orD

isch

arge

CD

UC

DU

or

not

BL

UE

Mor

eL

ab?

BL

UE

Doc

tor

Eva

luat

ion

1B

LU

E L

ab?

BL

UE

Mid

Ass

essm

ent

BL

UE

Bed

Ass

ignm

ent

BL

UE

Bed

Del

ay

PAC

E D

isch

arge

PAC

E B

edC

lean

ing

PAC

E D

ispo

sal

Wal

kin

Dis

posa

lW

alki

n D

octo

rE

valu

atio

nW

alki

n D

isch

arge

Wal

kin

Bed

Cle

ansi

ng

Wal

kink

Tri

age

Figu

re3(

a):

This

Real

Opt

flow

char

tre

pres

ents

asi

mpl

ified

EDw

orkfl

owat

Grad

y.

ED disposition to departure (Oct 1 31, 2009)

Figure 3(b): (Color online) The figure shows discharge destinations forOctober 2009 ED patients and the time taken from ED disposition to actualdeparture for each destination. Destinations are the telemedicine sign-up, intensive care unit, floor (normally staffed inpatient unit), isolationunit, and the stepdown unit (provides intermediate care between the ICUand floor).

Note that until the patient is admitted elsewhere, EDresources (in particular, the patient’s bed) are not freeto assign to a new patient. This information forms partof our RealOpt model for systems optimization.

Within RealOpt, optimization can be performed toensure the best operations and system performance(e.g., throughput, wait time, queue length, utilization).The resource allocation is modeled via a nonlinearmixed-integer program (NMIP). Resource examplesinclude labor, equipment, and beds. Constraints in themodel include (1) maximum limits on wait time andqueue length, which are dictated by the capacity of thewaiting room in most EDs, and the desire to quicklyservice patients; (2) range of utilization desired ateach station; (3) for each resource group, assignabil-ity and availability of resource types at each station(i.e., the skill set and the numbers of skilled personnelavailable); and (4) maximum limit on the cycle timeof the individual (i.e., ED LOS). The system param-eters in the simulation (i.e., queues, wait time, uti-lization, cycle time, and throughput) are performancevariables in the optimization. Because some functionsin the objective and constraints are not necessarilyexpressible in closed form, the problem is intractableby commercial systems. RealOpt is designed to over-come such computational bottlenecks by interweavingrapid system simulation and optimization (Lee et al.2013a, b). The appendix includes further details onthe model.

Dow

nloa

ded

from

info

rms.

org

by [

74.2

17.1

96.2

2] o

n 09

Aug

ust 2

016,

at 0

8:08

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


As we describe previously, machine learning is usedto identify discriminatory attributes that can predictwhether a patient will return to the ED. Within thesimulation, an individual patient is simulated thor-oughly, including medical conditions, arrival times,zones visited, and treatments received. Hence, in addi-tion to modeling the hospital operations, it also char-acterizes each individual by disease type, risk fac-tors, demographics, and payer types—knowledge thatthe machine-learning analysis uncovers. Upon com-pletion of a patient’s treatment and before discharge,the machine-learning classification rule is used withinthe simulation environment to predict whether thatpatient will return. If it predicts that the patient willreturn, it triggers an alarm; a nurse then determines ifthe patient should be sent to the clinical decision unitfor observation.

The novelty herein is the incorporation of patientcharacteristics and care patterns that the machine-learning framework uncovers within the RealOptsimulation-optimization environment. Hence, agents(representing patients) within the simulation presentdisease symptoms that challenge the care providers.They mimic the behavior of returning patients forwhom certain symptoms may not have been diag-nosed properly during previous ED visits. Further, themodel captures more than 200 processes, including EDconnected environments (e.g., discharge destinationsand factors external to ED), which contribute to delays

Phase I: Train: Aug. 2008–Feb. 2009 Phase II: Train: Oct.–Dec. 2010Validate: March–May 2009 Validate: Jan.–March 2011

Hospital statistics Simulated values Hospital statistics Simulated values

LOS Patient LOS Patient LOS Patient LOS PatientED zone (hours) volume∗ (hours) volume (hours) volume (hours) volume

Overall 10059 8,274 10049 8,446 7097 8,421 8002 8,398Blue zone 14054 2,141 13090 2,137 11040 2,107 11078 2,126Red zone 12054 2,097 11096 2,140 8098 2,083 8037 2,133Trauma center∗∗ 7085 271 7098 251 6080 268 6086 259Detention 13085 437 12093 407 10090 441 10053 432PACe 7090 2,037 8060 1,983 5010 1,920 5060 1,983Walk-in 3020 990 3030 992 2050 950 2088 940

Table 2: This table shows actual and simulated 30-day average LOS and throughput at Grady Hospital.Note. Remainder patients: ∗301 of the patients in this column include those who left before being seen,transferred to another facility, or provided no information. ∗∗These are airlift level 1 trauma patients. Gradytreats roughly 1,542 trauma patients per month; many enter through the ED and are treated in the blue zone.

in the ED workflow. The optimization component con-nects the multiple-resource allocation, as we describeearlier, with process and operations optimization overthe entire ED process network. The multiple-objectivefunction values are evaluated through the simulationprocess.

Model ValidationUsing the data collected, we simulated the hospitalenvironment and operations, and validated the sim-ulation results against an independent set of threemonths of hospital data. The model returned ED LOS,throughput, wait time, queue length, and other sys-tem statistics that are useful for performance measure-ment and comparison. For brevity, Table 2 includesonly LOS and throughput comparisons. The simula-tion results accurately reflect the existing ED systemperformance, with outcome metrics and performancestatistics consistent with their actual hospital values.

The average characteristics of Grady’s ED patientsdiffer markedly from national averages, especiallybecause so few have private insurance. In 2009 atGrady, only eight percent of the ED patients hadprivate insurance; more than 50 percent self-paidfor the service, and Medicaid and Medicare paidfor 36 percent. In contrast, nationally, approximately50 percent of ED patients have private insurance.Moreover, Grady is burdened by return visits fromuninsured individuals who use the ED as their pri-mary care facility. Table 3(a) shows Grady’s ED

Dow

nloa

ded

from

info

rms.

org

by [

74.2

17.1

96.2

2] o

n 09

Aug

ust 2

016,

at 0

8:08

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Nov.–Dec. 2009 72-hour return 30-day return

No. of No. of Percentage No. of PercentageAcuity level visits revisits of revisits revisits revisits

Total 15,168 824 5043 31279 210621: Immediate 367 17 4063 56 150262: Emergent 2,793 157 5062 651 230313: Urgent 6,595 385 5084 11531 230214: Less urgent 3,310 147 4044 651 190675: Nonurgent 1,531 90 5088 294 19020None—missing 572 28 4090 96 16078

Table 3(a): ED readmission statistics for the period from November toDecember 2009 are shown.

72-hour return 30-day return

10-fold Blind- 10-fold Blind-cross- prediction cross- prediction

Acuity level validation (%) accuracy (%) validation (%) accuracy (%)

1: Immediate 8309 8207 7803 75042: Emergent 7000 7000 7907 79003: Urgent 7001 7005 7805 78054: Less urgent 7101 7001 8002 80005: Nonurgent 7005 7005 7700 7805None—missing 7503 7407 8908 9101Overall 7100 7101 7903 7807

Payment type

Private 8605 8509 8407 8408insurance

Self-pay 6701 6703 7609 7606Medicare 7001 7009 7705 7709Medicaid 6601 6704 7605 7607

Table 3(b): The table illustrates 10-fold cross-validation results andblind-prediction accuracy for 72-hour and 30-day returns. The percent-age represents the percentage of patients with correct predictions.

readmission statistics for November to December2009, which were close to the national average.

Our goal in predicting readmissions is twofold:(1) capture the characteristics of the disease andtreatment patterns of readmitted patients to incorpo-rate their behavior within the simulation-optimizationenvironment; and (2) provide real-time guidance toED providers to identify individuals (for observation)before discharge to mitigate the number of avoidablereadmissions. Reducing the number of readmissionsimproves quality of care and provides financial andresource savings.

We apply the machine-learning framework usinga training set of 42,456 patients, and blind pre-dict using an independent set of 18,464 patients to

gauge the predictive accuracies. Table 3(b) summa-rizes the results. We select those results in whichboth the specificity and sensitivity are above 70 per-cent. Note that for self-pay and Medicaid patients,the accuracy is below 70 percent. We also observethat predicting insured individuals yields the high-est accuracy, because insured individuals use the EDonly when necessary. Obtaining high prediction accu-racy for patients who are not privately insured (e.g.,self-pay, Medicaid, Medicare) is difficult. We also notethat for 72-hour returns, prediction accuracy is high-est for acuity-level 1 patients, because their symp-toms and conditions are generally more conspicuous;in addition, 72-hour returns and 30-day returns showvariations.

Computational Results,Implementation, and EDPerformance Comparison

Phase I: ResultsWe performed systems optimization of the overall EDprocesses. In addition to the ED processes, the sys-tem model included other units in which ED patientsare being discharged, for example, the ICU, stepdown,floor, isolation unit, and telemedicine sign-up. InTable 4(a), we summarize the operational performanceaccording to improvement options using LOS andthroughput. When we optimized over the existing EDlayout, the system returned a global solution, whichcomprises Options 1–4. When we relaxed the layoutrestriction and optimized, it returned Options 1, 2,and 5 as the solution. Although these global solu-tions include a collection of changes and recommenda-tions that together result in the best overall operationsimprovement, we split the solution into individualoptions and individually simulated the effects of thesechanges to allow for prioritization and selection byhospital management for implementation.

Specifically, we separated the global solution intofive options according to change potential, and ana-lyzed the anticipated ED operations improvement.Next, we describe the five options and their predictedimpact.

Option 1. Combining registration and triage de-creases the LOS of blue- and red-zone patients by morethan one hour, with more significant gains by the most

Dow

nloa

ded

from

info

rms.

org

by [

74.2

17.1

96.2

2] o

n 09

Aug

ust 2

016,

at 0

8:08

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Actual hospitaloperations Simulation systems performance

March–May 2009 Systems improvement

Simulation Option 2: Option 3: Option 4: Option 5:output Option 1: Reduce Optimize Optimize Combine

Actual (using Aug.–Dec. System Combine lab/X-ray staffing in staffing in blue and red zoneshospital 2008 observed solution registration turnaround blue and red triage, walk in, with optimizedstatistics data for training) (Options 1–4) and triage (−15 min) zones and PACe staffing

OverallPatient volume 8,274 8,446 8,413 8,433 8,324 8,392 8,401 8,331LOS (h) 10059 10049 7033 10002 9022 9084 9049 7068Average total wait time (h) 4051 4034 1039 3095 2050 3087 3064 1076

Blue zonePatient volume 2,141 2,137 2,135 2,138 2,139 2,145 2,139 4,273LOS (h) 14054 1309 11008 12089 11083 13038 14000 8070

Red zonePatient volume 2,097 2,140 2,129 2,142 2,137 2,145 2,140 See aboveLOS (h) 12054 11096 8064 11034 10034 10062 12001 See above

TraumaPatient volume 271 251 251 250 249 251 251 271LOS (h) 7085 7098 6094 7051 7049 7074 7098 7070

DetentionPatient volume 437 407 410 411 410 408 411 401LOS (h) 13085 12093 10017 13095 11036 12046 13095 9016

PACePatient volume 2,037 1,983 1,970 1,988 1,966 2,001 1,979 1,989LOS (h) 7090 8060 3064 8060 7095 7074 4003 6063

Walk inPatient volume 990 992 989 997 996 990 998 971LOS (h) 3020 3030 109 3031 2086 302 2049 2094

Table 4(a): The table shows ED LOS and throughput comparisons for various systems improvement strategies.

severe (i.e., blue zone) patients. Detention patientsare registered separately; thus, they do not benefitfrom the change. We find no change in how traumapatients are admitted, and marginal improvement forless urgent patients.

Option 2. Reducing laboratory and X-ray turn-around time (by 15 minutes) drastically reduces blue-zone, red-zone, and detention patients’ LOS by morethan two hours. These savings are realized because59 percent of these patients require one laboratoryorder and 40 percent require two orders. The gain isalso realized for trauma patients, although to a lesserextent. PACe and walk-in patients seldom requirelaboratory or X-ray orders. The time reduction isachieved via bin-tracking on orders and improvedscheduled pickup and delivery between the ED andthe laboratory.

Option 3. Optimizing staffing in blue and red zonesreduces the LOS of blue- and red-zone patients by

more than one hour, with more significant reductionsobserved by red-zone patients, because nurses origi-nally operated at about 80 percent capacity in the bluezone and at 91 percent in the red zone. Detention-patient LOS also decreases because of using blue-zoneresources.

Option 4. Optimizing staffing in triage, walk-in, andPACe areas reduces LOS by about 30 minutes for blue-and red-zone patients; as expected, it has a majorimpact on PACe and walk-in patients, reducing LOSby 3.8 hours (−49 percent) and 42 minutes (−22 per-cent), respectively.

Option 5. Combining blue- and red-zone layoutswith optimized staffing offers substantial operationalefficiency. Before the ESI was introduced, patientswere sent to each color zone for similar complaintsand (or) severity, as set forth by hospital personnel,to streamline the treatment process, to be assignedto appropriate providers, or to anticipate complexity

Dow

nloa

ded

from

info

rms.

org

by [

74.2

17.1

96.2

2] o

n 09

Aug

ust 2

016,

at 0

8:08

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


of treatment. This also made revisits easier becausepatients would recall their previously assigned colorzone. With the establishment of the ESI and sophisti-cated triage, patients are assigned an acuity level toassist in the treatment process. The color zones nolonger serve their original purpose, although the hos-pital retains them (and appropriately uses them toaccommodate the ESI). At Grady, the blue and redzones are adjacent to each other and share the samelabor resources. Providers spend a good part of eachday walking back and forth between these two zonestending to patients. Our combined layout with opti-mized staffing provides operational efficiency, becauseit reduces LOS by more than five and three hoursfor the blue and red zones, respectively, and reducesmore than 40 percent of blue-zone LOS and 30 percentof red-zone LOS. Detention patients use blue-zoneresources and achieve a LOS reduction of about 26 per-cent. As expected, LOS for trauma patients improvesonly slightly.

In addition to the systems optimization, our time-motion studies and machine-learning analysis also ledus to make the following recommendation to hospitalmanagement.

Option 6. Allocate a separate area for walk-inpatients to be assigned a bed instead of at the ambu-lance triage area.

Option 7. Eliminate batching patients from thewalk-in area to a zone or PACe. Instead of accumulat-ing enough patients and taking a group of them to azone or PACe, service each patient based on his (her)arrival time.

Option 8. Eliminate batch discharges. Dischargepaperwork is performed for each patient wheneverthat patient is ready for discharge, rather than dis-charging them in groups.

Option 9. Create a clinical decision unit to observepatients before formal discharge to reduce avoidablereadmissions as a result of insufficient care, discrepan-cies in diagnosis, or premature discharge. This optionarises from the machine-learning analysis that pre-dicts patients who would be readmitted; providersthen observe them to mitigate the readmission prob-ability. This area is created by system optimization,which repurposes three beds from the blue zone andfour beds from the red zone. Since 2003, Grady had anobservation unit with six beds to manage ED patients

for whom extra time is required to determine dis-charge or hospital admission. The repurposed sevenbeds increase Grady’s ED observation capacity.

Option 10. Redirect nonurgent or walk-in patients toan alternative care facility.

Phase I: Adoption and ImplementationGrady management adopted Options 1–4, 7, and 8for implementation, but made a minor alteration toOption 1—it combined the registration and triage onlyat the ambulance arrival area. These changes, whichrequired no extra resources, were implemented byJuly 2009.

At that time, Option 9 was under discussion forimplementation because we recommended reallocat-ing or reoptimizing existing resources (i.e., labor,space, equipment). Option 10 was under considerationto raise funds to help pay for establishing the alterna-tive care facility.

Subsequently, based on follow-up time-motionstudies and independent best-practice benchmarkingtools that Grady employs (e.g., CMS core measures,National Association of Public Hospitals (NAPH)quality indicators, Press Ganey, and Leapfrog qual-ity indicators), the changes implemented by July 2009led to a LOS reduction of about three hours (frommore than 10 hours to slightly more than seven hours),as Table 4(b) shows for the Phase I adoption andimplementation. In January 2011, the hospital imple-mented the recommended clinical decision unit forobservation, using the machine-learning prediction totrigger the targeted treated ED patients for observa-tion. Figure 4 shows the actual reduction of 72-hourand 30-day return patients. For acuity levels 1 and 2,72-hour returns decreased by more than 30 percentand 7 percent, respectively. For 30-day returns, thereductions for these two levels were 24 percent and9 percent, respectively. Our Grady ED transformationwas timely. As a result of requirements in the Afford-able Care Act, the hospital does not receive paymentfor return visits; in addition, it must pay a penalty.Hence, reducing avoidable readmissions representsimproved care quality and provides financial savings.

These improvements raised confidence in our rec-ommendations, and prompted the hospital to use $1million of a donation from Kaiser Permanente to acton Option 10 of our recommendations—to open an

Dow

nloa

ded

from

info

rms.

org

by [

74.2

17.1

96.2

2] o

n 09

Aug

ust 2

016,

at 0

8:08

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Phase I: Comparison of ED performance (actual hospital monthly statistics)

Implementation of recommendations

Options 1–4, 7–9 Options 1–4, 7–10(clinical decision unit (redirect nonurgent visits

Original Options 1–4, 7, 8 for observation) to walk-in center)

March–May 2009 July 2009–Dec. 2010 Jan.–Aug. 2011 Sept.–Dec. 2011

Patient Reduction in Patient Reduction in Patient Reduction in PatientED zone LOS (l∗5 (h) volume LOS (l − l∗5 (h) volume LOS (l − l∗5 (h) volume LOS (l − l∗5 (h) volume

Overall 10059 8,274 −3000 8,395 −2086 8,421 −2029 8,364Blue zone 14054 2,141 −3026 2,525 −3014 2,317 −3022 2,603a

Red zone 12054 2,097 −3078 2,109 −3080 2,230 −3060 2,254Trauma center 7085 271 −1001 252 −1019 283 −1022 305a

Detention 13085 437 −3012 420 −2095 446 −3001 445PACe 7090 2,037 −3002 2,104 −3018 2,098 −3060 2,083Walk-in 3020 990 −100 945 −0085 970 −102 510b

Table 4(b): The table shows 30-day average LOS and throughput performance following the initial Phase Iimplementation.

aThe new trauma center was opened in November 2011.bA significant number of nonurgent ED patients have been redirected to the new walk-in center since

August 19, 2011, thus resulting in a significant decrease in ED walk-in patients.

alternative care facility, a walk-in center for low-acuitypatients. This facility opened in August 2011 (Williams2011). With confidence in improved ED efficiency, inOctober 2011, Grady also unveiled the Marcus TraumaCenter, which increases the number of trauma bedsfrom four to 15 (PRNewswire 2014). Based on the

0.0

5.0

10.0

15.0

20.0

25.0

Figure 4: (Color online) The graph compares the percentage of ED revisitsin 2010 and 2011. Note the significant reduction in 72-hour and 30-dayreturns following the installation of the clinical decision unit in 2011.

improved ED efficiency, our study recommended onlyone additional attending physician.

Phase II ResultsIn conjunction with the walk-in-center option and theincrease of beds in the trauma center from four to 15,the hospital gained an attending physician; however,the ED demand also increased. The hospital observeda slight increase in LOS from 7.9 hours to more than 8hours. Understanding that system improvement is anon-going effort in aligning demand with resources, theteam embarked on Phase II of the system optimizationeffort using existing resources.

In the following summary, we omit the performancereport for the PACe and walk-in, because the systemoutput from the various strategies offers only marginalLOS differences compared to the larger improvementsobserved from Phase I.

Table 5(a) summarizes the anticipated results basedon simulation and optimization. Specifically, globallyoptimizing the system resulted in an overall LOSreduction of 90 minutes. This entails global resourceallocation and changes in ED layout. The improve-ment is considerable, with major LOS reductions inthe blue and red zones (44 percent and 30 percent,respectively). Although the trauma center significantly

Dow

nloa

ded

from

info

rms.

org

by [

74.2

17.1

96.2

2] o

n 09

Aug

ust 2

016,

at 0

8:08

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Simulation systems performance

Actual hospital Global strategy: Option 11: Option 12:operations System optimization Optimize worker Combine blue and

Aug.–Dec. 2011 (resource + layout) allocation red zones

Overall LOS (hours) 8030 6.79 (−1051) 7.21 (−1009) 6.94 (−1036)Blue zone (hours) 11032 6.24 (−5008) 6.66 (−4066) 6.61 (−4071)Red zone (hours) 8094 6.24 (−2070) 6.19 (−2075) 6.61 (−2033)Trauma center (hours) 6063 6.46 (−0017) 6.16 (−0047) 6.47 (−0016)

Table 5(a): Phase II comparisons of potential ED performance show efficiency improvements using differentstrategies.

Phase II: Comparison of ED performance (actual hospital monthly statistics)

Implementation of Phase II recommendations

Original (from Phase I improvement) Option 11 (optimizing overall ED staffing)

Sept.–Dec. 2011 2012 Jan.–Dec. 2013

ED Zone LOS (l∗∗5 Reduction in LOS (l − l∗∗5 Reduction in LOS (l − l∗∗5

Patient volume 8,364 8,920 9,060Overall LOS 8030 h −1000 h −1016 hBlue zone 11032 h −3095 h −4005 h (−36%)Red zone 8094 h −2070 h −2052 h (−30%)Trauma center 6063 h −0035 h −0030 h (−5%)

Table 5(b): The 30-day average LOS and throughput performance improved as a result of the Phase IIimplementation.

increased its bed capability, it added only one attend-ing physician. As a result, trauma LOS improved byonly about 10 minutes, because the new facility had asignificant increase in trauma patients. Nevertheless,for trauma patients, particularly those who suffer fromtraumatic brain injury, 10 minutes can have a tremen-dous impact on outcome (e.g., survival, disablement,death), and is vital to the survival and quality of lifeof these patients.

Splitting the global strategies into Option 11 (opti-mal staffing) and Option 12 (optimal layout) resultedin similar minor LOS improvements in traumapatients. However, blue- and red-zone patients contin-ued to enjoy significant LOS reductions.

Phase II: Adoption and ImplementationTable 5(b) contrasts the ED performance before andafter the Phase II Option 11 implementation. Usingexisting resources and facility layout, Grady gainedefficiency and timeliness of care by simply globallyoptimizing resources across the ED. The net LOSreduction of four hours for high-severity patients (i.e.,

blue zone) is substantial and could translate to betterquality of care and outcomes. Even minor improve-ments in timeliness of care for trauma patients couldmake a difference between life and death, and have asignificant impact on quality of life for these patients.

Combining blue and red zones is viable because allpatients entering either zone require a consultationand generally require multiple resources or extensivediagnostic testing. However, such layout redesign maynot be desirable, because commingling patients withdifferent acuity levels may have detrimental effectson the treatment process; for example, care providersmay not be as focused. The net gain of combining thezones, even with optimizing resource usage across allareas, is less substantial for trauma patients. The hos-pital executives carefully weighed this option, and arenow confident that combining the zones will have anoverall positive impact. At the time of this writing, thehospital has received $77 million of sponsored fundingand has embarked on the ED facility layout redesign.

Dow

nloa

ded

from

info

rms.

org

by [

74.2

17.1

96.2

2] o

n 09

Aug

ust 2

016,

at 0

8:08

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


0.00

1.00

2.00

3.00

4.00

5.00

6.00

(%)

7.00

1 2 3 4 5

72-hour return 20102011

20122013

0.00

5.00

10.00

15.00

20.00

25.00

1 2 3 4 5

30-day return

(%)

20102011

20122013

Figure 5: (Color online) The graphs compare the percentage of ED revisits in 2010 (no intervention), 2011(Phase I), and 2012–2013 (Phase II). Note the significant drop in 72-hour and 30-day returns following thePhase I implementation. The machine-learning tool learns from revisit patterns and improves progressivelyas it adapts through the years. The level 4 and 5 patients who use the ED as their primary care service (i.e.,super utilizers) remain a challenge, especially for 72-hour returns.

To monitor the performance of the clinical decisionunit, Figure 5 contrasts the 72-hour and 30-day returnperformance for 2010–2013. Between Phase I and II,the 30-day return reduction shows substantial gain,especially among severe-acuity patients. Nonurgentpatients (level 5) return to EDs at high rates because itis often their only means of access to healthcare. Suchnonurgent readmission is unavoidable because somepatients come in with unrelated health complaints. Incontrast, although level 1 patients demand the mosturgent care, their diagnosis is typically very specific;upon discharge, they are well counseled with regardto follow-up care with their primary care providers,resulting in lower returns to the ED. Mid-level acuitypatients have higher rates of return because of the less-specific nature of their complaints and (or) diagnosis.

Figure 6 shows LOS trends for the ED zones throughthe various stages of implementation. Specifically,adoption of the initial optimization of overall staffingand process consolidation significantly reduced LOSacross all zones (from the first to the second bar). Thisimplementation did not require additional resourcesor financial investment. When the clinical decisionunit was established in 2011, Grady experienced amarginal increase in LOS across all zones, becausesome patients were selected for observation to reducepotential returns (third bar). This also very slightlyaffected the LOS of the blue and red zones, because

space and labor resources were repurposed for theclinical decision unit. In September 2011, the alter-native walk-in center was opened, drastically reduc-ing LOS for ED walk-in patients (fourth bar). Thedifference in LOS across other zones was marginal;however, overall LOS increased slightly because the

2012

AC

2013

Figure 6: (Color online) The graph compares LOS from 2009 to thepresent. Grady has sustained a steady ED LOS since the 2009 systemimprovement implementation. The graph shows: March–May 2009 (orig-inal performance, first bar), July 2009–Dec. 2010 (after Phase I imple-mentation, second bar), Jan.–Aug. 2011 (after implementation of theclinical decision unit for observation, third bar), Sept.–Dec. 2011 (afterimplementation of the walk-in center to redirect nonurgent patients,and expansion of the trauma center, fourth bar), and 2012–2013 (afterPhase II implementation, fifth and sixth bars). The overall average LOSin 2013 was 7.14 hours.

Dow

nloa

ded

from

info

rms.

org

by [

74.2

17.1

96.2

2] o

n 09

Aug

ust 2

016,

at 0

8:08

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


number of walk-in patients decreased significantly.Although the throughput in the ED and trauma cen-ter increased steadily over the years (by approximately16.2 percent), the LOS from 2012 to the present stayedclose to constant, indicating that the earlier improve-ment was being sustained. The clinical decision unitreduced potential avoidable returns, thus helping tosave hospital resources, reduce penalties, and improvethe quality of patient care. Since November 2013, theunit has expanded to 15 beds. The walk-in centerhas helped to relieve Grady’s large healthcare bur-den of Medicaid and Medicare patients. By redirectingnonurgent ED patients, the hospital saved valuableresources and reduced the costs needed to unnecessar-ily treat these patients in the ED. This has also reducedthe number of patients who leave without being seenby more than 32 percent.

Benefits and ImpactsThis OR analytical work and the subsequent imple-mentation and successes are extremely important toGrady. As a safety net healthcare provider, Gradymust make transformative changes to improve effi-ciencies and reduce expenses so that it can continueto provide care to a significant segment of the pop-ulation that is underserved medically. The goal ofour work is to significantly improve the efficiencyand timely delivery of quality care to Grady’s EDpatients. In the opinion of Grady executives and med-ical staff members, our OR analytical work madepossible and substantially facilitated the benefits andimpacts listed next.

Quantitative BenefitsOur work has improved the timeliness of emergencycare. From the beginning of Phase I to the present,

LOS for ED patients LOS for ED patientsHospital discharged to hospital (hrs) discharged home (hrs) Average ED LOS

LAC/USC (Los Angeles) 1708 607Cook County (Chicago) 1500 600

8–11 hrsParkland (Dallas) 1101 503Grady (2014) 903 607 7.1 hrsGrady (2008) before improvement 1305 1000 10.6 hrs

Table 6: In this table, we compare the LOS in major safety net hospitals (http://www.Hospitalcompare.hhs.gov).

the overall average LOS decreased by 33 percent(10.59 hours to 7.14 hours), while average total wait-ing time decreased by 70 percent. This contrasts withan ED LOS of 8 to 11 hours in comparable safetynet hospitals (see Table 6). The reductions are mostsignificant for high-acuity patients: LOS decreasedby more than 50 percent for both the blue andred zones (−7027 hours and −6028 hours, respec-tively); the LOS in the trauma zone decreased by20 percent (−1052 hours). Next, we list quantitativeimprovements.

Improved Efficiency of Emergency Care. Facili-tated by the creation of the walk-in center, the im-provements allowed Grady to increase its ED annualthroughput (i.e., number of patients treated) by morethan 7.8 percent (+81114), its trauma volume by8.4 percent (+11664), and its volume of severe traumacases (i.e., patients facing life-and-death situations) by14 percent (+417), and reduce the number of patientswho leave without being seen by more than 30 per-cent (−51553). Moreover, it made these improvementswithout increasing its ED staff or facilities. The use ofthe clinical decision unit decreased avoidable 72-hourand 30-day readmissions among the acuity-levels 1–3patients by 28 percent (−602). This produced directfinancial and resource savings for the hospital andhad a positive impact on patient-care quality mea-sures. The alternative walk-in center serviced morethan 32 percent of the nonurgent ED cases outside theED treatment area, thus reducing the hospital’s finan-cial burden (by treating these patients in a lower-costarea) and ensuring proper ED resource usage.

Annual Financial Savings and Revenues. From2008 to 2012, the reduction in revisits resulted in $7.5million of savings in penalties. The walk-in center for

Dow

nloa

ded

from

info

rms.

org

by [

74.2

17.1

96.2

2] o

n 09

Aug

ust 2

016,

at 0

8:08

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


nonurgent conditions reduced ED costs by $21.6 mil-lion and resulted in $12.5 million in additional rev-enue. ED and trauma efficiency increased the revenueby $96.6 million. Expansion of trauma care resulted in$51.8 million in revenue. For a critical safety net hos-pital with $1.5 billion of annual economic impact, onlyeight percent of which is paid by private insurance, the$190 million financial gains have a tremendous impacton maintaining Grady’s financial health.

The ED, often called the front door to a hospital,serves as a source of hospital inpatient admissions,which on average generate more revenue than ED-only admissions. For Grady Hospital, the ED providesabout 75 percent of inpatient admissions. Thus, theED’s increased throughput and other improvementsplayed a major role in the significant revenue increasesshown in Grady’s annual financial reports.

Encouragement of External Sponsorship. In part,as a result of the rigorous OR-driven recommenda-tions, Grady has been able to document success intimeliness of care and operational efficiencies, thusfacilitating increased philanthropic donations. TheKaiser Foundation contributed $1 million (Williams2011) to establish an alternative care site (walk-incenter) for low-acuity patients. A $20 million gift fromthe Marcus Foundation (PRNewswire 2014) enabledGrady to create a world-class stroke and neurosciencecenter and a state-of-the-art trauma center. The ORadvances and subsequent ED transformation giveinvestors confidence in sponsoring projects that willbenefit the hospital and its patients.

Qualitative BenefitsOur work has saved lives and reduced morbidityand disabilities. Efficient ED operations allow theED to more quickly treat patients with time-sensitiveconditions. The shortened LOS demonstrates that

ED service 2013 volume Increase in volume Reduced death and disability

Airlift trauma patients 3,395 patients 417 patients (+14%) All need immediate carefor life-and-death situations

Trauma patients 15,992 1,665 (+804%) Death ∼ 56Disability ∼ 160

Comprehensive care 39,059 2,001 (+5052%) Disability ∼ 390Extended care 29,645 902 (+209%) Disability ∼ 296

Table 7: These estimates of potential death and disability reductions resulted from increasing patient volumesat Grady in 2013. Volume increases shown compare 2013 and 2012.

patients move from the ED and receive appropri-ate care in the appropriate setting in a timelier man-ner. For high-acuity patients, quicker response dur-ing the golden hour of treatment (i.e., a period ofabout one hour following traumatic injury) duringwhich prompt medical treatment will likely preventdeath can mean the difference between life and death,disability, or returning to a normal life. Faster door-to-computerized-tomography (CT) scan and door-to-tissue-plasminogen-activator (a clot-dissolving drug)administration for stroke patients, and faster door-to-antibiotics for pneumonia patients have decreasedlong-term disability and death. More acute traumapatients can be treated, saving more lives. Improvedtimeliness and service quality directly translate toimproved quality of life for patients and decreasedmorbidity and mortality, and make a difference inwhether a patient is treated and released, or is admit-ted to the hospital (see Table 7).

Health Cost Reductions. ED timeliness and effi-ciency of care have a broad impact on patient qual-ity of life and healthcare spending. Timeliness andimproved quality of care improve outcomes, and con-sequently lead to indirect savings of hundreds of mil-lions of dollars in ongoing care and managementof patients. In addition, reducing disability allowspatients to lead normal lives. The estimate of the valueof one life in the United States is $50,000–$100,000 peryear of life saved (Owens 1998). We emphasize thatalthough quality and systems efficiency have been ourfocus, the monetary savings (for both the providersand the patients) are real and critical to our nationalhealthcare system. This is especially true for safety nethospitals such as Grady, which feel the burden moreacutely than other hospitals, given that many of itsservice costs are not reimbursed.

Dow

nloa

ded

from

info

rms.

org

by [

74.2

17.1

96.2

2] o

n 09

Aug

ust 2

016,

at 0

8:08

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Continuous Improvement and Adaptive Advances.The hospital has been able to achieve its targetedgoal of ED LOS of seven hours and sustain overallimprovement for over five years. With ED demandcontinuing to grow, maintaining a culture of con-tinuous improvement is key to sustaining goodperformance.

Improved Quality of Care in Other Facilities. Themodel can be generalized and has been tested andsuccessfully implemented in 10 other EDs. The bene-fits across these EDs are consistent with the substan-tial benefits Grady achieved. The ED volumes at these10 sites range from 30,000 to 80,000 patients per year.Upon implementation, they have experienced a totalthroughput increase from 15 to 35 percent, a reductionof revisits of severe acute patients from 19 to 41 per-cent, a LOS reduction from 15 to 38 percent, and areduction in the number of patients who leave withoutbeing treated from 35 to 50 percent.

Grady has applied the technology in other units,including medication error analysis for the pharmacyand hospital-acquired conditions (HACs) in the ED,operating room, and intensive care units. HAC isone of the 10 major causes of death in the UnitedStates. Our surgical site infection (SSI) study at Gradyinvolved reducing mediastinitis after cardiac surgery.Nationwide, the 700,000 open-heart surgeries per-formed each year result in infection rates of 0.5–5 per-cent. Of those infected, the mortality rate is 40 percent.On average, an additional 30 days of hospital LOSand (or) one extra surgical procedure are required.The SSI rate at Grady was 23 percent in 2010. Theteam implemented transformative changes, includingstrategic preoperative procedures for both inpatientsand outpatients and optimal timing and dosing of pre-operative antibiotics in July 2011. The infection ratedecreased to 1.5 percent between July 2011 and Jan-uary 2012, and has been zero percent since Febru-ary 2012. The team is now conducting a study on jointsurgeries, bloodstream infection, and catheter-inducedurinary-tract infection.

Scientific AdvancesThe collaborative effort between hospital researchersand OR scientists resulted in scientific advances ontwo fronts, as follows.

Hospital Care Delivery Advances. The new systemcouples machine learning, and simulation and opti-mization decision support to improve the efficiencyand timeliness of care in the ED, while reducing avoid-able readmissions. The model allows a hospital toglobally optimize its ED workflow, taking into accountthe uncertainty of human disease characteristics andcare patterns, to drive the patient LOS and wait time toa minimum. It provides a comprehensive analysis ofthe entire patient flow from registration to discharge,and enables a decision maker to understand the com-plexities and interdependencies of individual steps inthe process sequence; ultimately, it allows a hospitalto perform systems optimization to achieve the mostoptimum performance.

The model focuses on system optimization thatresults in improvements in LOS and waiting timethrough resource allocation, system consolidation, andoperations optimization without attempting to changethe behavior of healthcare providers or patients.Rather, the system captures the human behavior andoptimizes the workflow process to achieve optimalresults. Although changing human behavior can resultin significant gains, we understand that such changesmay be more costly in terms of training and alter-ing habits; this is particularly true for teaching hos-pitals at which rotations of residents and healthcaretrainees are common. The potential to introduce newerrors also exists. Thus, we accept the variability inhuman behavior and services and incorporate theseelements into our model to reflect workflow andhuman characteristics.

OR Advances. The novelty of our OR-drivenanalytical work includes performing systems opti-mization within the ED simulation environment;incorporating treatment patterns and patient char-acteristics dynamically and stochastically within theED operations and quality-improvement framework;modeling ED readmission using—simultaneously—demographics, socioeconomic status, clinical infor-mation, hospital operations, and disease behavioralpatterns; modeling ED interdependencies involvingother hospital units; and integrating a machine-learning framework within the simulation-optimiza-tion environment.

We acknowledge the computational challenges ofsuch large-scale complex models in data collection for

Dow

nloa

ded

from

info

rms.

org

by [

74.2

17.1

96.2

2] o

n 09

Aug

ust 2

016,

at 0

8:08

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


model validation, parameter estimation, and globalsystem optimization. The machine-learning frame-work and the DAMIP model have been provento be NP-hard (Brooks and Lee 2014); hence, theyrequire both theoretical and computational break-throughs (Brooks and Lee 2010, 2014; Lee 2007; Leeand Maheshwary 2013; Lee and Wu 2007; Lee et al.2003, 2014). However, once the predictive rule hasbeen established, it can analyze and predict patientreturn patterns in nanoseconds, opening up real-time target patient intervention. We derived poly-hedral theory and applied it to the solution strate-gies for DAMIP (Brooks and Lee 2010, 2014; Lee andMaheshwary 2013; Lee et al. 2014).

Because of the complexity of simultaneously sim-ulating dynamic system behavior and optimizingoperational performance, solving within our simula-tion-optimization framework remains a challenge. Wecaution that our solutions, although obtained rapidly,are not proven to be optimal. Nevertheless, ourinvestigations indicate that the solutions are close tooptimal.

Future PlansWe acknowledge the proactive attitude of our health-care collaborators who worked tirelessly to learn oursystem advances and present our recommendations tokey stakeholders for change transformation to makethis project a success. Challenges remain: the team willclosely monitor the facility-layout redesign, and mea-sure important outcome metrics to gauge its impacton overall performance and patient care. In addition,important regulatory compliances and critical and del-icate issues related to mental health treatment mustbe addressed. We will also address the super utilizers,those underinsured patients who utilize the ED withup to 33 visits per month. Further, beyond the day-to-day operations, we will carry out strategic plan-ning. The system developed allows healthcare man-agers to adapt the model as the ED environment andoperations evolve. Moreover, work is underway toapply the system to improve efficiency and quality inother units. Such a dynamic and flexible computer-ized system is critical for sustained continuous opera-tional improvement. It will help Grady adapt quicklyto changes in the healthcare environment, and ensure

its survival as an essential safety net hospital servingthe Atlanta community and beyond.

Related photos, presentations, hospital-insidernotes, and an Institute of Medicine/National Acad-emy of Engineering letter concerning the significanceof the work are available at http://www2.isye.gatech.edu/medicalor/EDadvances.

AcknowledgmentsThis study is partially supported by funding from theNational Science Foundation to the first author [GrantsIIP-0832390 and IIP-1361532]. The authors would like tothank C. Allen, C. Girard, S. Lahlou, D. Meagh, J. Phillips,A. Widmaier, R. L. Zhou, and H. Z. Zhang from GeorgiaTech for performing the time-motion study on this project.Special thanks go to Grady’s caretakers: Deborah Western,Manuel Patterson, Nadia Ralliford, Jill Cuestas, and all thenurses for their support of this project and for helping withdata collection and interviews.

AppendixOptimization-Based Classifier: Discriminant Analysis viaMixed-Integer ProgramIn this section, we briefly describe DAMIP. Suppose we haven entities (e.g., patients) from K groups (e.g., returning ornonreturning) with m features. Let G = 81121 0 0 0 1K9 be thegroup index set, O= 81121 0 0 0 1n9 be the entity index set, andF= 81121 0 0 0 1m9 be the feature index set. Also, let Ok, k ∈Gand Ok ⊆ O, be the entity set that belongs to group k. More-over, let Fj , j ∈F, be the domain of feature j , which couldbe the space of real, integer, or binary values. The ith entity,i ∈ O, is represented as 4yi1xi5 = 4yi1xi11 0 0 0 1 xim5 ∈ G×F1 ×

· · ·×Fm, where yi is the group to which entity i belongs, and4xi11 0 0 0 1 xim5 is the feature vector of entity i. The classifica-tion model finds a function ë : 4F1 × · · · ×Fm5→G to clas-sify entities into groups based on a selected set of features.

Let �k be the prior probability of a randomly chosenentity being in group k and fk4x5 be the group condi-tional probability density function for the entity x ∈ �m ofgroup k, k ∈G. Also let nh denote the number of entitiesfrom group h1 and �hk ∈ 40115, h1k ∈G, h 6= k1 be the upperbound for the misclassification percentage that group h enti-ties are misclassified into group k. DAMIP seeks a parti-tion 8P01P11 0 0 0 1 PK9 of �K , where Pk, k ∈G is the region forgroup k, and P0 is the reserved-judgment region with enti-ties for which group assignment are reserved (for potentialfurther exploration).

Let uki be the binary variable to denote if entity i is classi-fied to group k. Mathematically, DAMIP can be formulatedas (DAMIP) (Lee 2007, Lee et al. 2003).

max∑

i∈O

uyi i(D1)

s.t. Lki =�kfk4xi5−∑

h∈G1h6=k

fh4xi5�hk1

∀ i ∈ O1 k ∈G1 (D2)

Dow

nloa

ded

from

info

rms.

org

by [

74.2

17.1

96.2

2] o

n 09

Aug

ust 2

016,

at 0

8:08

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


uki =

{

1 if k = arg max801Lhi2 h ∈G91

0 otherwise1

∀ i ∈ O1 k ∈ 809∪G1 (D3)∑

k∈809∪G

uki = 11 ∀ i ∈ O1 (D4)

∑

i2 i∈Oh

uki ≤ ��hknh�1 ∀h1 k ∈G1 h 6= k1 (D5)

uki ∈ 801191 ∀ i ∈ O1 k ∈ 809∪G1

Lki unrestricted in sign1 ∀ i ∈ O1 k ∈G1

�hk ≥ 01 ∀h1 k ∈G1 h 6= k0

The objective function (D1) maximizes the number ofentities classified into the correct group. Constraints (D2)and (D3) govern the placement of an entity into each ofthe groups in G or the reserved-judgment region. Thus, thevariables Lki and �hk provide the shape of the partition ofthe groups in the G space. Constraint (D4) ensures thatan entity is assigned to exactly one group. Constraint (D5)allows the users to preset the desirable misclassification lev-els, which can be specified as overall errors for each group,pairwise errors, or overall errors for all groups together.With the reserved judgment in place, the mathematical sys-tem ensures that a solution that satisfies the preset errorsalways exists.

Mathematically, we have proven that DAMIP is NP-hardand that the resulting classification rule is strongly univer-sally consistent, given that the Bayes optimal rule for clas-sification is known (Brooks and Lee 2010, 2014). Compu-tationally, DAMIP is the first multiple-group classificationmodel that includes a reserved judgment and the ability toconstrain the misclassification rates simultaneously withinthe model. Furthermore, we have demonstrated in real-world applications that DAMIP works well by using auniform prior probability and a normal group conditionalprobability density function. They serve to transform theattributes from their original space to the group space(Brooks and Lee 2010, 2014; Koczor et al. 2013; Lee 2007;Lee and Wu 2007; Lee et al. 2003, 2012a; Nakaya et al. 2011;Sturm et al. 2010). In Brooks and Lee (2010, 2014), we haveshown that DAMIP is difficult to solve, and we applied thehypergraphic structures that Lee and Maheshwary (2013)and Lee et al. (2014) derived to efficiently solve theseinstances. Empirically, DAMIP can handle imbalanced datawell; thus, it is suitable for the ED readmission analysis,when compared against other classification approaches (Leeet al. 2012a).

The predictive model maximizes the number of cor-rectly classified cases; therefore, it is robust and not skewedby errors committed by observation values. The associ-ated optimal decision variable values (Lki and �hk5 formthe classification rule, which consists of the discriminatoryattributes; examples include patient chief complaint, diag-nosis, whether IV antibiotics were ordered, trainee and (or)

resident involved, primary nurse, time when the patientreceived an ED bed to time until first medical doctorarrived. Using this rule, blind prediction of whether a newpatient will return to the ED can be performed in real time.

One can use alternative objectives, for example, by plac-ing different weights on each group. In Lee (2007), we dis-cuss various alternative objectives that take into account dif-ferences in the relative cost of different types of classificationerrors. We tested these alternatives on hospital readmissionstudies (Lee et al. 2012a). For the paper herein, we reportthe best combination when applied to the Grady data.

Nonlinear Mixed-Integer Program forMultiple-Resource AllocationThe nonlinear mixed-integer program (MIP) used in allo-cating resources was built on top of the nonlinear MIPformulated in Lee et al. (2009) and solved using the simula-tion-optimization framework described in Lee et al. (2013a).In all three papers (including this work), an initial solutionis obtained via a fluid model as a warm start to a resourceoptimization model; the results of the optimization areentered into a simulation model that estimates the system’saverage wait time, queue length, and utilization. If the solu-tion satisfies all the input ranges, this solution is returned.Otherwise, the system looks for violated constraints (fromamong wait time, queue length, and cycle time), and deter-mines violated service blocks. Service blocks are all the ser-vices and (or) processes that a patient might undergo inan ED visit, including triage, registration, PACe examina-tion, walk-in, and laboratory tests. Once these blocks areidentified, optimization will be performed on them. Thistime, however, the objective is to minimize a total violationpenalty. The process continues until convergence.

There are two key distinctions between this work andthe earlier work by Lee (e.g., Lee et al. 2012b, 2013a).First, machine learning that predicts patient pattern andtreatment characteristics is integrated into the simulation-optimization decision framework. Second, this work modelsmultiple resource groups at each station. The optimizationwithin each simulation step maximizes throughput with-out exceeding the existing resources, and includes weightson how to use staff skills. For example, if a nurse and aphysician can both perform a specific task, it may be moreexpensive to use a physician than a nurse (or vice versa).The user ranks them in the input, and we use these ranksas weights in the optimization process.

References

Allaudeen N, Vidyarthi A, Maselli J, Auerbach A (2011a) Redefin-ing readmission risk factors for general medicine patients.J. Hospital Medicine 6(2):54–60.

Allaudeen N, Schnipper JL, Orav EJ, Wachter RM, Vidyarthi AR(2011b) Inability of providers to predict unplanned readmis-sions. J. General Internal Medicine 26(7):771–776.

Dow

nloa

ded

from

info

rms.

org

by [

74.2

17.1

96.2

2] o

n 09

Aug

ust 2

016,

at 0

8:08

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


American College of Emergency Physicians (2014) National ReportCard. Press release: Nation’s grade drops to a dismal D+ forfailure to support emergency patients. Accessed September 7,2014, http://www.emreportcard.org/Content.aspx?id=534.

American Trauma Society (2014) Trauma center levels explained.Accessed August 18, 2014, http://www.amtrauma.org/?page=TraumaLevels.

Ashby M, Ferrin D, Miller M, Shahi N (2008) Discrete event simu-lation: Optimizing patient flow and redesign in a replacementfacility. Accessed September 7, 2014, http://www.informs-sim.org/wsc08papers/200.pdf.

Bahensky JA, Roe J, Bolton R (2005) Lean sigma-will it work forhealthcare? J. Healthcare Information Management 19(1):39–44.

Benbassat J, Taragin M (2000) Hospital readmissions as a measureof quality of health care: Advantages and limitations. Arch.Internal Medicine 160(8):1074–1081.

Billings J, Dixon J, Mijanovich T, Wennberg D (2006) Case findingfor patients at risk of readmission to hospital: Developmentof algorithm to identify high risk patients. BMJ 333(7563):327.

Brooks JP, Lee EK (2010) Analysis of the consistency of a mixedinteger programming-based multi-category constrained dis-criminant model. Ann. Oper. Res. Data Mining 174(1):147–168.

Brooks JP, Lee EK (2014) Solving a multigroup mixed-integer pro-gramming-based constrained discrimination model. INFORMSJ. Comput. 26(3):567–585.

Centers for Disease Control (2010) National hospital ambulatorymedical care survey: 2010 emergency department summarytables. Accessed August 20, 2014, http://www.cdc.gov/nchs/data/ahcd/nhamcs_emergency/2010_ed_web_tables.pdf.

Davidson G, Moscovice I, Remus D (2007) Hospital size, uncer-tainty and pay-for-performance. Health Care Financing Rev.29(1):45–57.

DeFlitch C, Eitel D, Geeting G, Cherukuri R, Escott M, Smith J,Arnold J, Paul SA (2007) Physician directed queuing (PDQ)improves health care delivery in the ED: Early results. Ann.Emergency Medicine 50(3):S125–S126.

de Moura HC (2007) Financial crisis at Atlanta’s Grady Hos-pital. Accessed September 20, 2014, http://www.gpb.org/healthdesk/financial-crisis-at-atlantas-grady-hospital.

Derlet RW (2002) Overcrowding in emergency departments: In-creased demand and decreased capacity. Ann. EmergencyMedicine 39(4):430–432.

Derlet RW, Richards JR (2000) Overcrowding in the nation’s emer-gency departments: Complex causes and disturbing effects.Ann. Emergency Medicine 35(1):63–68.

Derlet RW, Richards JR (2002) Emergency department overcrowd-ing in Florida, New York, and Texas. Southern Medical J. 95(8):846–849.

Derlet RW, Richards JR, Kravitz R (2001) Frequent overcrowdingin US emergency departments. Acad. Emergency Medicine 8(2):151–155.

Eitel DR, Rudkin SE, Malvehy MA, Killeen JP, Pines JM (2010)Improving service quality by understanding emergencydepartment flow: A white paper and position statement pre-pared for the American Academy of Emergency Medicine.J. Emergency Medicine 38(1):70–79.

Hagen MS, Jopling JK, Buchman TG, Lee EK (2013) Priority queu-ing models for hospital intensive care units and the impactsto severe case patients. Accessed October 1, 2014, http://www.ncbi.nlm.nih.gov/pmc/articles/PMC3900220.

Hasan O, Meltzer DO, Shaykevich Bell CM, Kaboli PJ, AuerbachAD, Wetterneck TB, Arora VM, Zhang J, Schnipper JL (2010)Hospital readmission in general medicine patients: A predic-tion model. J. General Internal Medicine 25(3):211–219.

Joynt KE, Orav EJ, Jha AK (2011) Thirty-day readmission rates forMedicare beneficiaries by race and site of care. J. Amer. MedicalAssoc. 305(7):675–681.

Kansagara D, Englander H, Salanitro A, Kagen D, Theobald C,Freeman M, Kripalani S (2011) Risk prediction models for hos-pital readmission. J. Amer. Medical Assoc. 306(15):1688–1698.

Kirby SE, Dennis SM, Jayasinghe UW, Harris MF (2010) Patientrelated factors in frequent readmissions: The influence ofcondition, access to services and patient choice. AccessedDecember 1, 2014, http://www.biomedcentral.com/content/pdf/1472-6963-10-216.pdf.

Koczor CA, Lee EK, Torres RA, Boyd A, Vega JD, Uppal K, YuanF, Fields EJ, Samarel AM, Lewis W (2013) Detection of differ-entially methylated gene promoters in failing and nonfailinghuman left ventricle myocardium using computational analy-sis. Physiol. Genomics 45(14):597–605.

Kolker A (2008) Process modeling of emergency departmentpatient flow: Effect of patient length of stay on ED diversion.J. Medical Systems 32(5):389–401.

Lee EK (2007) Large-scale optimization-based classification mod-els in medicine and biology. Ann. Biomedical Engrg. 35(6):1095–1109.

Lee EK, Maheshwary S (2013) Facets of conflict hypergraphs. SIAMJ. Optim. Forthcoming.

Lee EK, Wu TL (2007) Classification and disease predictionvia mathematical programming. Seref O, Kundakcioglu OE,Pardalos PM, eds. Data Mining, Systems Analysis and Optimiza-tion in Biomedicine (American Institute of Physics, Melville,NY), 1–42.

Lee EK, Gallagher RJ, Patterson DA (2003) A linear programmingapproach to discriminant analysis with a reserved-judgmentregion. INFORMS J. Comput. 15(1):23–41.

Lee EK, Pietz F, Benecke B (2013b) Service networks for publichealth and medical preparedness: Medical countermeasuresdispensing and large-scale disaster relief efforts. Herrmann, J,ed. Handbook of Operations Research for Homeland Security(Springer, Berlin), 167–196.

Lee EK, Surana V, Shapoval A (2014) Facets for generalized conflicthypergraphs. SIAM J. Optim. Submitted.

Lee EK, Chen C-H, Pietz F, Benecke B (2009) Modeling andoptimizing the public-health infrastructure for emergencyresponse. Interfaces 39(5):476–490.

Lee EK, Maheshwary S, Mason J, Glisson W (2006a) Large-scale dispensing for emergency response to bioterrorism andinfectious-disease outbreak. Interfaces 36(6):591–607.

Lee EK, Maheshwary S, Mason J, Glisson W (2006b) Decisionsupport system for mass dispensing of medications for infec-tious disease outbreaks and bioterrorist attacks. Ann. Oper. Res.148(1):25–53.

Lee EK, Yang AY, Pietz F, Benecke B (2011) Public health and med-ical preparedness. Yih Y, ed. Handbook of Healthcare DeliverySystems, Chap. 41 (CRC Press, Boca Raton, FL), 41-1–41-40.

Lee EK, Pietz F, Benecke B, Mason J, Burel G (2013a) Advanc-ing public health and medical preparedness with operationsresearch. Interfaces 43(1):79–98.

Lee EK, Yuan F, Hirsh DA, Mallory MD, Simon HK (2012a)A clinical decision tool for predicting patient care character-istics: Patients returning within 72 hours in the emergencydepartment. Accessed October 1, 2014, http://www.ncbi.nlm.nih.gov/pmc/articles/PMC3540516.

Lee EK, Chen C-H, Brown N, Handy J, Desiderio A, Lopez R,Davis B (2012b) Designing guest flow and operations logisticsfor the dolphin tales. Interfaces 42(5):492–506.

Dow

nloa

ded

from

info

rms.

org

by [

74.2

17.1

96.2

2] o

n 09

Aug

ust 2

016,

at 0

8:08

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Lewin Group (2002) Emergency department overload: A grow-ing crisis. Accessed October 1, 2014, http://www.aha.org/content/00-10/EdoCrisisSlides.pdf.

McCarthy ML, Zeger SL, Ding R, Levin SR, Desmond JS, Lee J,Aronsky D (2009) Crowding delays treatment and lengthensemergency department length of stay, even among high-acuitypatients. Ann. Emergency Medicine 54(4):492–503, e4.

Medeiros DJ, Swenson E, DeFlitch C (2008) Improving patient flowin a hospital emergency department. Mason S, Hill R, RoseO, Moünch L, eds. Proc. 2008 Simulation Conf. (Institute ofElectrical and Electronics Engineers, Washington, DC), 1526–1531.

Minott J (2008) Reducing hospital readmissions. Accessed Octo-ber 1, 2014, http://www.academyhealth.org/files/publications/ReducingHospitalReadmissions.pdf.

Moskop JC, Sklar DP, Geiderman JM, Schears RM, Bookman KJ(2009) Emergency department crowding, part 2—Barriers toreform and strategies to overcome them. Ann. EmergencyMedicine 53(5):612–617.

Nakaya HI, Wrammert J, Lee EK, Racioppi L, Marie-Kunze S,Haining WN, Means AR, et al. (2011) Systems biology ofseasonal influenza vaccination in humans. Nature Immunology12(8):786–795.

Nugus P, Holdgate A, Fry M, Forero R, McCarthy S, Braithwarte J(2011) Work pressure and patient flow management in theemergency department: Findings from an ethnographic study.Academic Emerging Medicine 18(10):1045–1052.

Osei-Anto A, Joshi M, Audet AM, Berman A, Jencks S (2010)Health care leader action guide to reduce avoidable readmis-sions. Accessed October 1, 2014, http://www.hret.org/care/projects/resources/readmissions_cp.pdf.

Owens DK (1998) Interpretation of cost-effectiveness analyses.J. General Internal Medicine 13(10):716–717.

Popovich MA, Boyd C, Dachenhaus T, Kusler D (2012) Improv-ing stable patient flow through the emergency departmentby utilizing evidence-based practice: One hospital’s journey.J. Emergency Nursing 38(5):474–478.

PRNewswire (2014) The Marcus foundation gives $20 millionto Grady Health System. Accessed October 1, 2014, http://www.prnewswire.com/news-releases/the-marcus-foundation-gives-20-million-to-grady-health-system-61849312.html.

RealOpt (2012) RealOpt user manual. Accessed December 1, 2012,http://www2.isye.gatech.edu/medicalor/realopt/.

Richardson LD, Hwang U (2001) America’s health care safetynet: Intact or unraveling? Academic Emergency Medicine 8(11):1056–1063.

Scuteri J, Fodero L, Pearse J (2011) Determining a threshold hos-pital size for the application of activity-based funding. BMCHealth Services Res. 10(Supp.1):A10.

Southern DA, Quan HG, William A (2004) Comparison of the Elix-hauser and Charlson/Deyo methods of comorbidity measure-ment in administrative data. Medical Care 42(4):355–360.

Sturm JJ, Hirsh DA, Lee EK, Massey R, Weselman B, Simon HK(2010) Practice characteristics that influence nonurgent pedi-atric emergency department utilization. Academic Pediatrics10(1):70–74.

Taylor TB (2001) Threats to the health care safety net. AcademicEmergency Medicine 8(11):1080–1087.

United States Department of Health and Human Services (2014)A policy framework for targeting financially vulnerable safetynet hospitals. Accessed August 18, 2014, http://aspe.hhs.gov/health/reports/02/dsh/ch2.htm.

VanSuch M, Naessens JM, Stroebel RJ, Huddleston JM, Williams AR(2006) Effect of discharge instructions on readmission of hospi-talised patients with heart failure: Do all of the joint commis-sion on accreditation of healthcare organizations heart failurecore measures reflect better care? Quality Safety Health Care15(5):414–417.

Westert GP, Lagoe RJ, Keskimäki I, Leyland A, Murphy M (2002)An international study of hospital readmissions and relatedutilization in Europe and the USA. Health Policy 61(3):269–278.

Wiler JL, Gentle C, Halfpenny JM, Heins A, Mehrotra A, MikhailMG, Fite D (2010) Optimizing emergency department front-end operation. Ann. Emergency Medicine 5(2):142–160.

Williams M (2011) Grady to open ER alternative offering lowercosts. Accessed October 1, 2014, http://www.ajc.com/news/news/local/grady-to-open-er-alternative-offering-lower-costs/nQKxb/.

Eva K. Lee is a professor in the H. Milton StewartSchool of Industrial and Systems Engineering at GeorgiaInstitute of Technology, and director of the Center forOperations Research in Medicine and HealthCare, a cen-ter established through funds from the National ScienceFoundation (NSF) and the Whitaker Foundation. The cen-ter focuses on biomedicine, public health, and defense,advancing domains from basic science to translational med-ical research; intelligent, quality, and cost-effective deliv-ery; and medical preparedness and protection of criticalinfrastructures. She is also the co-director of the Centerfor Health Organization Transformation, an NSF Indus-try/University Cooperative Research Center. Lee partnerswith hospital leaders to develop novel transformationalstrategies in delivery, quality, safety, operations efficiency,information management, change management, and orga-nizational learning. She is a distinguished scholar in healthsystems at the Health System Institute, a joint research insti-tute between Georgia Tech and Emory University Schoolof Medicine. Lee earned a PhD at Rice University in theDepartment of Computational and Applied Mathematics,and received her undergraduate degree in mathematics fromHong Kong Baptist University, where she graduated withhighest distinction. Her research focuses on mathematicalprogramming, information technology, and computationalalgorithms for risk assessment, decision making, predictiveanalytics and knowledge discovery, and systems optimiza-tion. She has made major contributions in advances to med-ical care and procedures, emergency response and medicalpreparedness, healthcare operations, and business opera-tions transformation.

Hany Y. Atallah is the chief of service for emergencymedicine at Grady Health System. He is an assistant pro-fessor in the Department of Emergency Medicine at EmoryUniversity School of Medicine. He manages the operationsof the North Georgia’s premier level 1 trauma center, whichincludes the largest emergency department in the south-east. He is responsible for all clinical, graduate medicaleducation and administrative activities occurring withinthe service. His primary responsibility is for clinical ser-vice performance—inclusive of clinical outcomes, physician

Dow

nloa

ded

from

info

rms.

org

by [

74.2

17.1

96.2

2] o

n 09

Aug

ust 2

016,

at 0

8:08

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


performance and professionalism. Dr. Atallah holds degreesfrom Washington University in St. Louis and a medicaldegree from New York Medical College. He is a memberof the Eastern Association for the Surgery of Trauma, theSociety for Simulation in Healthcare, and is a fellow of theAmerican College of Emergency Physicians.

Michael D. Wright is a senior healthcare executive with areputation for exceptional vision and operational leadershipin academic medical centers, teaching hospitals, faith-basedand community health systems. Wright served as senior vicepresident of operations at Grady Health System from 2010to 2013. Previously, Wright has served as chief operatingofficer of Provena Saint Joseph Hospital, a 200-bed commu-nity hospital in Elgin, Illinois; vice president of operationsat WakeMed Health and Hospitals, a 550-bed teaching hos-pital in Raleigh, North Carolina; and regional director ofoperations with Johns Hopkins Community Physicians, adivision of Johns Hopkins Medicine in Baltimore, Maryland.He is currently the president of a consulting firm specializ-ing in hospital operations consulting and interim leadershipmanagement. Wright received a bachelor’s degree in healthscience and policy from the University of Maryland Balti-more, and a master’s in business administration from theUniversity of Baltimore. He is a fellow in the American Col-lege of Healthcare Executives, and an active member of theNational Association of Health Services Executives.

Eleanor T. Post is a chief nursing officer at the Rock-dale Medical Center in Conyers, Georgia. Prior to joiningRockdale, she was the vice president of emergency depart-ment and trauma at Grady Memorial Hospital from 2009–2013. She has over 20 years of experience in leadership anddirector roles in healthcare, leading units including patientcare services, critical care and trauma, adult services, nurs-ing, quality assurance, and clinical services. In these roles,she manages the strategic planning, development and oper-ations of the units and the overall workflow. At Grady,she was instrumental in the successful implementation ofan electronic medical record and tracking system. She hasincreased core measures in pneumonia, improved chargecapture in ED, decreased door-to-door time, decreased themedian length of stay, reduced employee turnover, andimproved customer satisfaction. Post received her master’sdegree in nursing from New York University, and her bach-elor’s in nursing from Wagner College in Staten Island,New York.

Calvin Thomas IV is vice president of the Health Divi-sion at Ivy Tech Community College in Indianapolis, Indi-ana, the largest community college in the country with 31campuses and over 200,000 students. Prior to his position

in education, Thomas has been a successful senior health-care executive in hospitals in four states (California, Florida,Georgia, and Missouri). In his various roles as a hos-pital executive, he has been responsible for the devel-opment of a $22 million orthopedic & spine institute, afull-service Hospitalist program, the onboarding electronichealth record from a paper process, and in one health-care system has grown shrinking revenue to as much as$500 million. He served as senior vice president of opera-tions at Grady Health System from 2008 to 2010. Thomasreceived his Bachelor of Science degree from Harris-Stowe State University in healthcare management and amaster’s degree in healthcare leadership from DartmouthCollege.

Daniel T. Wu is the chief medical information officerof the Grady Health System and also serves as the asso-ciate medical director of the Grady Emergency Department.He is an associate professor in the Department of Emer-gency Medicine at Emory University School of Medicine.He has been instrumental in Grady’s recent transformationfrom a paper-based patient record to an integrated electronichealth record for which the institution has achieved HIMSSStage 6. He also assists with the daily operations of northGeorgia’s premier level 1 trauma center, which includes oneof the largest emergency departments in the southeast. Wugraduated from the University of Chicago and obtained hismedical degree from Columbia University College of Physi-cians and Surgeons. He completed an internship in internalmedicine at St. Vincent’s Hospital in New York City and hisemergency medicine residency at Cook County Hospital inChicago.

Leon L. Haley Jr. is the Emory executive associate dean,clinical services at Grady and chief medical officer of theEmory Medical Care Foundation. He is an associate pro-fessor in the Department of Emergency Medicine at Emoryand formerly served as deputy senior vice president ofmedical affairs, chief of emergency medicine for the GradyHealth System, and vice chairman of the Emory Depart-ment of Emergency Medicine. Dr. Haley holds degreesfrom Brown University and the Universities of Pittsburghand Michigan. Dr. Haley is board certified in emergencymedicine and a fellow of the American College of Emer-gency Physicians. He is a member of the American Collegeof Emergency Physicians, American College of HealthcareExecutives, and American College of Physician Executives.He was a member of the IOM Committee on Health andInsurance Status. Dr. Haley has received research fundingfrom DoD, SAMSA, RWJ, and the Healthcare Foundation ofGeorgia.D

ownl

oade

d fr

om in

form

s.or

g by

[74

.217

.196

.22]

on

09 A

ugus

t 201

6, a

t 08:

08 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.

Health & Medicine

Machine learning, simulation and optimization in healthcare