Predicting Hospital Capacity and Efficiency

James P. McGlothlin

, Sriveni Vedire

, Hari Srinivasan

, Amar Madugula

Srinivasan Rajagopalan

and Latifur Khan

Fusion Consulting Inc, Irving, TX, U.S.A.

University of Texas at Dallas, Richardson, TX, U.S.A.

Keywords: Predictive Analytics, Data Warehousing, Patient Movement, Discrete Event Simulation.

Abstract: Hospitals and healthcare systems are challenged to service the growing healthcare needs of the population

with limited resources and tightly restrained finances. The best healthcare organizations constantly seek

performance improvement by adjusting both resources and processes. However, there are endless options

and possibilities for how to invest and adapt, and it is a formidable challenge to choose the right ones. The

challenge is that each potential change can have far reaching effects. This challenge is exacerbated even

further because it can be very expensive for a hospital to experience logjams in patient movement. Each and

every change has a “ripple” effect across the system and traditional analytics cannot calculate all the

ramifications and opportunities associated with such changes. This project uses historical records of patient

treatment plans in combination with a virtual discrete event simulation model to evaluate and predict capacity

and efficiency when resources are added, reduced or reallocated. The model assigns assets as needed to

execute the treatment plan, and calculates resulting volumes, length of stay, wait times, cost. This provides a

valuable resource to operations management and allows the hospital to invest and allocate resources in ways

that maximize financial benefit and quality of patient care.

1 INTRODUCTION

Hospitals and healthcare systems, especially US-

based academic healthcare institutions, are under

constant pressure to streamline and achieve more with

limited resources and finances. With the growth of

electronic medical record systems (EMRs), health-

care data warehouses, and business intelligence

technologies, there is more information available to

identify problems and opportunities for improvement.

For example, we have robust dashboards which show

length of stay for patient cohorts and how these values

compare to benchmarks. We have information

around patient time in “boarder” status, i.e. being

boarded in one department when they belong in

another. Examples include patients in the emergency

department (ED) who have been admitted to the

hospital, patients in surgical recovery rooms who are

no longer under the effects of anesthesia, and patients

in intensive care units (ICUs) who no longer require

critical care. We have information from patient-

reported data, clinical engineering interfaces, staff

time and attendance, and patient surveys. There is a

staggering and growing amount of data available.

The challenge though is to use all this information

to make good decisions and initiate valuable change.

We can create change by adding or reallocating

resources and by changing processes. Resources that

could be adjusted include staff and staff schedules,

beds and bed accommodations, operating rooms

(ORs) and OR allocation schedules, imaging

resources, clinic and ambulatory surgery locations,

staffing and hours and more. Example processes that

can be manipulated include: house and bed

management; ED registration, triage and rooming

procedures; preventive care initiatives; isolation

process; discharge procedures; care management

protocols; and urgent care facilities. It is challenging

enough to predict the cost and return on investment

(ROI) from such changes to make educated decisions.

However, what really makes the challenge difficult is

that each change can have far-reaching effects. For

example, adding ED resources could create a logjam

of ED patients waiting for an inpatient bed, or cause

significant congestion in inpatient departments.

Adding an OR room or adjusting operating schedules

can create problems placing patients after

perioperative recovery. Even a simple change like

562

McGlothlin, J., Vedire, S., Srinivasan, H., Madugula, A., Rajagopalan, S. and Khan, L.

Predicting Hospital Capacity and Efﬁciency.

DOI: 10.5220/0006658905620570

In Proceedings of the 11th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2018) - Volume 5: HEALTHINF, pages 562-570

ISBN: 978-989-758-281-3

adjusting discharge order times can create a logjam of

patients waiting for the orders to be processed. The

challenge is exacerbated because it can be very

expensive for a hospital to experience logjams in

patient movement. For example, an admitted patient

who is still in an ED bed is acquiring additional cost

without compensation, and very likely getting

reduced care. A patient occupying a regular bed, but

requiring critical care treatment, is not only receiving

reduced treatment, but is also creating additional staff

requirements for the non-critical care department.

Even small changes such as housekeeping procedures

or staffing can have far-reaching effects across the

hospital system.

In addition to the operational challenges above,

there are both opportunities and risks for the actual

clinical treatment of the patient. It is important to

ascertain that any change in process or staffing does

not reduce care. Generally, we measure quality of

care by key performance indicators (KPIs) including

readmission rate, mortality rate and rate of hospital-

acquired infections. Additionally, disposition should

be included to indicate that a discharge to home is a

better outcome than dispositions such as discharge to

skilled nursing facility. Finally, patient experience

(customer satisfaction) results can be included.

Additionally, treatment processes can be

improved. These include disease-specific clinical

protocols such as initiatives for sepsis, exacerbated

chronic obstructive pulmonary disease (COPD),

congestive heart failure, pneumonia and stroke.

There are also initiatives around specific clinical

events such as ventilation, blood transfusions

(McGlothlin 2017), administration of broad spectrum

anti-biotics, management of central lines and

catheters. Each of these can have far-reaching

effects. For example, improving ventilation protocols

can reduce reintubation and instances of ventilator-

acquired pneumonia, and can result in reduced

readmissions to ICU units. A significantly successful

clinical program can increase the stress on the

hospital by increasing the survival rate and putting

pressure on lower acuity units to absorb more

patients.

There is simply no meaningful way for anyone to

manually determine all the possible ramifications,

both positive and negative, from a change in resource

allocation, process or treatment. In this project, we

are proposing a digital simulation model which

coupled with bio-statistics can provide much more

meaningful and actionable insight.

2 BACKGROUND

In our investigation and design of this project we

looked at literature and industry use cases

surrounding:

 Predictive analytics in hospitals

 Tools for what-if scenarios for healthcare

models

 Predictions models in other industries

Discrete event simulation is the process of

creating a digital model of specific events and rules.

Every treatment or movement in a hospital can be

considered an event. Tools and systems for discrete

event simulation are sometimes referred to as “digital

twins”. Digital twins create a complete virtual model

of a physical asset or process, and can be networked

together. Discrete event simulation and digital twins

are not new technology. They have been used for

years to simulate events that are too hard to test

manually (Fishman 1978), such as the impact of heat

or wind speed on a plane (Tuegel 2011). Also, they

are used to simulate entire systems which are

connected by well-defined rules. For example, they

could be used in automobile manufacturing to

determine if there is a cost advantage to a more

expensive paint which dries more quickly (Grieves

2014). This would be based on the capacity of the

additional manufacturing systems to absorb

additional auto capacity more rapidly. There are

many commercial digital twin products available on

the market, as well as open source products such as

Ditto (Glocker 2017) and OMNet++ (Varga 2001)

and SimJava (Howell 1998).

Despite the long history of this technology, it has

become much more mainstream in recent years.

Digital twins were recently identified by Gartner as

one of the ten significant trends of 2017 (Gartner

2017). The reason for this sudden grown is IoT (the

internet of things). As there are more small devices

connected to the network, and so much is known

about these devices and sensors, there is more

opportunity for digital modelling.

Despite the growth of this technology, it has rarely

been applied to healthcare, which is a much less

predictable model. A patient treated quickly and

politely may still decide to leave the hospital against

medical advice. A healthy patient might suddenly

have a totally unexpected aneurism. A patient whose

life was just saved may still be bitter and unsatisfied.

Despite being optimally staffed for a normal Friday

night, any particular Friday night may have an

unforeseeable catastrophe such as a mass shooting or

an earth quake. System-wide predictive analytics in

Predicting Hospital Capacity and Efﬁciency

563

healthcare is always challenging and discrete event

simulation models are no exception.

The other reason IoT has accelerated the

emergence of digital twin technology is because it

provides real and discrete information. One of the

major drawbacks of the digital twin approach has

always been that it relied on the correct modelling of

the business process. Hospitals and clinicians are

often adapting to emergencies or unusual events and

standardized processes are hard to develop or

monitor. IoT allows us to do things like monitor

actual staff and patient movement, clinical

instrumentation and other true sources of data to infer

business processes and outliers which were

previously hard to realize.

In specific contexts, there has been significant

work on predictive analytics in healthcare. Sutter

Health was able to accurately predict 30 day

readmission rates and leverage the information to

target and reduce readmissions (Ng 2014) (Jamei

2017). Similarily, (Bardhan 2014) was able to predict

readmission for congestive heart failure after analysis

of data from 67 hospitals. Parkland Hospital has been

able to accurately predict daily census and several

hospitals have used similar technology to improve

coding and reimbursement (Bradley 2010). (Maguire

2013) describes a successful predictive data mining

project specific to diabetes. There are countless more

examples, and we have tried to leverage what worked

for them, but they tend to be limited to specific patient

cohorts and single data points. While digital models

for hospitals has been widely discussed, we were

unable to find a successful system-wide project in the

literature.

3 PROPOSED APPROACH

We will detail the implementation plan and phases for

our project in section 5 after describing our specific

requirements in section 4. In this section, we will just

give a quick overview of the approach.

The general approach is to:

1. Extract patient treatment plans from historical

encounters. These treatment plans must separate

what is necessary to heal the patient from what

happened in the historical encounter due to

hospital inefficiency or capacity. We use

biostatistical analysis to transform actual events

into a treatment plan for that patient.

2. Model the resources and allocation rules based

on the business processes of the hospital. This

process will include both interviewing subject

matter experts about the processes and profiling

the historical data to infer and validate the

proposed business rules.

3. Create a discrete event simulation model which

places random patients into the hospital

according to historic trends and then assigns

resources to the patients according to their

treatment plan

4. Collect statistics around the encounters as they

pass through the model hospital.

5. Repeat the randomization many times to

determine both the median values for length of

stay, volume and other KPIs, and also worst and

best case scenarios and the probability of

significant logjams.

Once this is done, this model will support adjusting

the resources and allocation rules, adjusting the

treatment plans, and adjusting the patient flow based

on market analysis and trends. Each of these creates a

what-if scenario analyzed with the same system

above, and multiple changes can be done in the model

together to evaluate an entire proposal.

4 REQUIREMENTS

Before we outline our solution and implementation

plan we need to more clearly identify the

requirements for the system.

4.1 Measures

The following list of measures and KPIs was

identified by our Operational Excellence team. The

goal is that the solution will be able to accurately

predict each of these measures.

▪ Throughput and Volume

» Length of stay (total and in each area)

» Volume (ED, OR, and transfer)

» External transfer acceptance rate

» Isolation patient days/hours

» Census by hour and type

» Discharges by hour

▪ Wait times

» Boarding time (ED, ICU and OR)

» Discharge order to discharge

▪ Utilization and Productivity

» OR utilization

» Staff productivity

» Ancillary utilization including imaging,

labs, pharmacy, telemetry

» Bed Utilization

▪ Finances

» Contribution margin

» Activity based costs

HEALTHINF 2018 - 11th International Conference on Health Informatics

564

» Cost per discharge

▪ Patient and clinical Metrics

» Readmission rate

» Mortality rate

» Wellness scores

» Patient experience scores

Each of these measures we will want to look at by

specific patient cohorts based on diagnosis, payor,

unit, day, time, location, procedure, age, etc.

4.2 Assets and Resources

The goal is that this project support what-if scenario

analysis based on adding or removing resources or

changing schedules. The following is the list of such

resources determined so far:

▪ Beds

There can only be one patient in a bed at a time.

Whether a bed is appropriate for a patient is

based on patient class (inpatient, outpatient,

observation), level of care (such as ICU,

intermediate, acute) and service (specialities

such as cardiology or neurology). ED beds

support combination of acuity and age.

▪ Rooms

Rooms include one or more beds. Multiple beds

in the room can only be filled if the genders and

isolation status of the patients matches. For

example, we cannot place a patient with

clostridium difficile colitis (C. diff) with a patient

who does not, without risking serious harm to the

other patient. Isolation is managed in our system

with specific orders.

▪ Staffing

Staffing includes nurse and physician staffing by

unit, anesthesia staff, pharmacy staff, imaging

staff including technician and radiologists,

housekeeping staff, transport staff, and ancillary

staff including labs and pharmacy. Staffing

volume can be adjusted by hour, unit or location.

▪ Operating Rooms

▪ OR hours

The operating rooms support various volumes of

staffed rooms at different times and days of the

week. Moreover, there are primetime hours set

by day of week and location. Non-emergent

surgeries are delayed until prime hours.

▪ OR service block allocation

Our hospital holds operating rooms for specific

services according to the service block

allocation. This allocation can be changed.

▪ Imaging resources

Imaging devices such as MRI machines can be

added, removed, or given a new schedule.

The way our system works is we first build the list of

resources currently in the hospital using our data

warehouse. Then we set up the rules for allocating

the resources by working with the subject matter

experts in the hospital. Finally, we use our historical

data to validate these business process definitions and

adjust as needed.

Once we have all of the resources and rules, we

can create our model to “lock” these resources by

assigning them to specific patient encounters and

treatment plans. The system closely correlates a

semaphore locking algorithm. Once the model is

built, the user will be able to add or delete resources

or to change the business rules defining which

encounters can use the resources.

4.3 Treatment or Process Adjustment

We would like our solution to support adjusting

specific treatment plans, clinical processes or

operation processes, in addition to adjusting

resources. For example, if we are implementing a

stroke initiative that has been shown to reduce stroke

length of stay by 10% at other hospitals, we can use

our model to predict our measures given that all

stroke patients need 10% less treatment time. If we

have an initiative to write discharge orders or lab

orders earlier in the day, we can analyze the potential

effect of this change, and the appropriate alteration in

staff schedules. If we have a program to reduce red

blood cell transfusions by 50%, we can analyze the

likely effect of this improvement on additional results

such as length of stay, readmissions, hospital-

acquired infection rate, cost, and lab productivity.

4.4 Use Case Examples

In this section we will list example use cases we have

been given as potential questions this solution would

be able to answer.

1. Volume changes. Do we have the capacity

necessary to achieve the annual projections for

growth by service line? Do we have the capacity

necessary to handle the increase expected from

expanded primary care coverage?

2. What-if scenarios:

a. Asset reallocation including bed

assignment, OR hours and staffing

b. Process changes such as increased

transfers to auxiliary hospitals or

reduction in outpatient usage of beds

c. Treatment changes such as reduced

readmissions or performing triage in

waiting rooms

Predicting Hospital Capacity and Efﬁciency

565

d. Performance improvement such as

reduced OR turnaround times

4.5 Goals

As we look at this vast number of potential measures,

use cases and what-if scenarios, it is important we

focus on the strategic goals. These goals are to

optimally assign resources and create processes that

increase volume and patient satisfaction while

reducing cost, length of stay, readmissions and

mortality.

5 IMPLEMENTATION PLAN

5.1 Phases

This project is broad so our goal is to create a model

that can be expanded throughout several phases.

Phase 1: Proof of Concept

In this phase, we will create the model specific only

to beds and surgeries. We will calculate the volume,

throughput and utilization measures excluding

staffing and block utilization. We will concentrate on

a single use case: What will happen when we

experience growth by service line according to

projections?

Phase 2: Adjusting assets and treatment plans

In this phase we will build on phase 1 by allowing

beds to be reassigned, operating rooms to be added

and treatment plans to be changed. We will focus on

the complex scenario: Predict the measures given that

we reassign X number of beds from the surgical ICU

to the medical ICU, we reduce surgical ICU length of

stay 10% and we open one OR on Saturdays for non-

emergent cases. This shows how scenarios need to

include multiple factors to truly allow what-if

analysis, and the ability of the model to support this.

Phase 3: Adjusting operating room schedules

Currently, some operating rooms are pre-assigned to

specific services on specific days. This is a virtual

floating OR and does not specify the physical room

where the surgery takes place. In this phase, we will

support what-if scenarios where the blocks are

reassigned and we will include the block schedules in

assigning OR rooms and staff to encounters. We will

also add the measure of OR block utilization.

Phase 4: Anesthesia and ancillary staffing

In this phase, we will add measures around staff

productivity, we will include staff capacity in our

model and we will support what-if scenarios of

changing staff levels specific for anesthesia, labs and

images. We are not including nursing or attending

physician staffing to reduce complexity. Nurses are

generally staffed to bed occupancy anyway. To add

staffing to our solution, we will bring in data from

human resources and from the time and attendance

tracking system.

Phase 5: Cost

In this phase we will add cost data to get an

understanding of the financial impact of the what-if

scenarios from the previous phases.

Future Work

In the future, we can see going beyond manual what-

if scenario configuration to an optimization model

where all possible changes are considered and

optimal reassignments are proposed by our AI engine.

5.2 Team

The best opportunity for success in this project comes

from building a diverse team of experts in different

complimentary disciplines. Our team includes data

warehouse architects, statisticians, academic experts,

operational leaders, clinicians and user interface

specialists. Our purpose in detailing the team breakup

is to explain that we do not view this as simply a

technology or IT project. The best opportunity for

success is to create a project directed by the business

which includes clinicians, researchers and academics,

and technologists.

5.3 Technology

We have investigated using both commercial and

open source digital twin products, but instead, our

plan involves leveraging the technology the hospital

already has invested in. This includes Microsoft SQL

Server and an advanced enterprise data warehouse

which will support both calculating the KPIs for the

current environment and providing the historical

patients and treatment plans needed for our model.

This tight integration allows us to continue to support

our enterprise data definitions and “single version of

the truth” in both the actual data warehouse and the

new virtual model. We also utilize Cisco DV, a data

virtualization platform that allows us to quickly

integrate multiple data sources and manage

information independent of physical source. This is

important as we bring in human resource data, cost

information and time and attendance records. Cisco

DV platform can also be used to enhance our system

with auxiliary data such as avoidable delays or

hospital-acquired infections from our epidemiology

HEALTHINF 2018 - 11th International Conference on Health Informatics

566

Figure 1: Data analytics system architecture.

system. Additionally, we are leveraging SAP HANA

(Farber 2012), an in-memory data analytics appliance

which will provide exceptional performance and has

a built in predictive analytics library. We are also

using R, an open source language specifically

targeted to advanced statistics and predictive

analytics. System R, originally a IBM product, is an

open source suite of such tools and algorithms.

“R has also been identified by the FDA as suitable

for interpreting data from clinical research.” (Smith

2012) Finally, we are using the dashboard

visualization tool Tableau. Tableau is already in place

at our hospital as the delivery mechanism for

dashboards and data discovery, and it supports both

HANA and R. This allows us to create a common set

of tools, data definitions and interfaces rather than

introducing new technology and learning curves.

Figure 1 shows our system architecture.

5.4 Specific Implementation Plan

In this section, we will detail our implementation

tasks and plan for successful phased execution of this

project.

5.4.1 Calculating the KPIs

For the first phase, we are targeting the KPIs around

discharge, transfer and surgical volume, and around

length of stay, census, bed utilization, isolation and

boarding times. Most of these data points are in the

enterprise data warehouse, but we need to aggregate

them appropriately and add a few fields around the

transfer center and isolation status. It is important that

we have the current and historical state of these

metrics calculated in a way that is accurate and uses

matching patient cohorts. This allows us both to

compare our what-if scenarios to current state, and,

very importantly, to test our predictive model.

5.4.2 Data Preparation

We need to create our historical patient encounters.

We do this as a multiple stage process. First, we

extract from our data warehouse and build an event

log for each hospital encounter. This log includes

timing information concerning all patient movement.

This will include arrival and triage time, each bed

movement, each order which changes the patient’s

status or level of care, each level of care or service

change, and the admit and discharge orders.

Additionaly,l we add events surrounding surgeries:

when they were scheduled, when they were

performed, how long they took, etc.

We then use our statistical tools and

biostatisticians expertise to convert our historical

event logs to treatment plans. For example, consider

a patient who spent 1 day in inpatient ICU status, but

physically still in the ED and two days in ICU status

in a ICU bed. Does their treatment require two days

of ICU care or three? The answer is probably

somewhere in between because they were receiving

care while in boarder status in the ED, but not the

same care as they would have received in the ICU.

The goal of the statistics in this phase is to determine

the relationship of boarder time to length of stay.

Predicting Hospital Capacity and Efﬁciency

567

Once we have this information, we store with each

historical encounter a mini treatment plan.

5.4.3 Modeling the Assets

For phase 1, the assets we care about are limited to

just beds and operating rooms. For each bed outside

of ED, we generate a bridge table which shows every

class/service/level-of-care combination supported by

that bed. We also map the beds to rooms (which is

already in the data warehouse). Now we have a list

of resources which we can use to support events on

our encounters and mini treatment plan.

It is also interesting to note that our hospital has

reassigned or added these bed resources in the past.

We load this information as type-2 historical

information using standard data warehousing

methodology. This allows us to view our historical

records in context and more accurately extract

appropriate treatment plans. Also, because we have

past change scenarios and actual results, we have real

and meaningful data and situations available for us to

validate our prediction model.

For ED, we have six types of beds: rapid intake,

adult, express care, pediatric, diagnostic and surgery.

ED is modelled somewhat differently. The only

deciding factor are acuity, age, and time of day

(express care is not open 24 hours a day).

For operating rooms, we load the list of operating

rooms/locations, the prime time hours and the staffed

hours for emergent cases. We are not concerned with

service block schedules in the first two phases.

5.4.4 Building the Allocation Agent

We build a set of coded procedures which support the

treatment plan of patient encounters by:

▪ Assigning the patient to an appropriate bed when

they are in boarder or waiting room status

▪ Adding a hold to the other bed in a room (based

on gender and isolation status)

▪ Moving patients around to support holds (for

example if two males are in separate rooms with

empty beds in a unit and a female needs a beds,

the males can be moved together)

▪ Assigning an operating room based upon the

schedule request and executing a surgery

5.4.5 Additional Statistics

We calculate additional statistics to support our

model. These include:

▪ Histograms which tell us how many patients

to expect at specific times

This allows us to more accurately feed

patients into our predictive model.

▪ Transfer center statistics to show us the

relationship between the wait time for

approval and the cancellation rate

This will allow us to predict transfer volume

and acceptance rate.

▪ ED statistics to show us the relationship

between delays in treatment and patients

leaving prematurely

This will allow us to predict how many

patients will leave prematurely based upon

the predicted waiting times.

▪ Discharge processing statistics to show us

the relationship between discharge order

times and volume to actual discharge times.

This will allow us to predict when a patient

will actually be discharged in relationship to

when the discharge order was written.

5.4.6 Virtual Model Execution

To execute our model, we place patients in the

hospital (via ED arrival, direct admit or transfer

request) according to our histograms from the

statistical analysis. We then loop every “15 minutes”

on our virtual clock and add new patients, discharge

patients, and attempt to move patients according to

their treatment plan. At any point in time, a patient is

either in appropriate status or boarder status. For

example if the treatment plan says the patient needs 3

days in ICU and they are currently assigned an ICU

bed and have only been in ICU one day in our virtual

system, they are appropriately placed and we will not

attempt to move them during this time loop. If they

are in intermediate level of care according to their

treatment team, but they are in that same ICU bed, we

will attempt to assign them an intermediate bed. If

we are able to do so, we will release the ICU bed.

Each time a patient leaves the hospital in our virtual

model, we record statistics about their encounter to

match our KPIs.

The priority order we assign patients to beds is

very important in our model execution. We will use

business rules that define which types of patients it is

most important to place first.

We load our patients completely randomly. We

take a number between 0 and 1 and multiply it by our

number of patient encounter treatment plans to

choose one. To make up for this randomization, we

run the same patients through with different random

variations thousands of times to calculate the median

KPI values and the histogram of each measure. This

allows us to not only predict the measures but also

HEALTHINF 2018 - 11th International Conference on Health Informatics

568

show where the potential for challenges or logjams is.

For example, it is possible that a given scenario

reduces length of stay on average, but has a greater

chance of increasing length of stay during stress

situations (high volume fluctuations).

5.4.7 What if Scenarios

The scenario we are supporting for Phase 1 simply

involves adjusting the patient encounters according to

given growth projections by service line. We can

achieve this with our statistics which define patient

encounters to add to the virtual hospital model.

For phase 2, we support three types of

adjustments: beds (add/delete/reallocate), operating

rooms (add/delete/change hours) and treatment plan.

Our system will allow the user to set up a

scenario. They can remove any bed in the system.

They can add a bed and define the class/level of

care/service combinations the new bed supports.

They can also specify rules for adjusting patient

treatment plans, such as reduce ICU length of stay for

all sepsis patients by 10%.

5.4.8 Testing

Before we use our model to make decisions, it is

obviously important to be certain our predictions are

accurate. To achieve this our plan is to use a

traditional training and testing process. We will train

using historical data from encounters 6-24 months

ago and then test by attempting to accurately predict

the last 6 months for the hospital.

Additionally we will use past changes for testing

purposes. Consider if 6 months ago our hospital

added 5 beds to the adult medical ICU. We could

train the system with encounters before 6 months ago,

then set up a what-if scenario where these 5 beds are

added. We use the predictive engine to predict the

resulting volumes and length of stay, and test these

predictions against our last 6 months of actual

hospital statistics.

6 CONCLUSIONS

We have proposed a staged solution to allow

hospitals to create “what-if scenarios” and predict

system results from such scenarios over a large

number of important measures. We have created an

achievable, targeted plan which delivers value in each

of several short phases as it builds the entire model.

Our model leverages the EMR and enterprise data

warehouse, well-known data mining techniques and

bio-statistical algorithms. It uses biostatistics,

discrete event simulation and business process

modelling in tandem. It is achievable, measurable

and flexible. We believe it will create a solid

foundation for predictive analytics for our hospital

system.

REFERENCES

Glocker, G., Jaekle, T., 2017. Eclipse Ditto. Viewed

August 2017. https://projects.eclipse.org/proposals/

eclipse-ditto.

Tuegel, E. J., Ingraffea, A. R., Eason, T. G. and

Spottswood, S. M., 2011. Reengineering aircraft

structural life prediction using a digital twin.

International Journal of Aerospace Engineering, 2011.

Grieves, M., 2014. Digital twin: Manufacturing excellence

through virtual factory replication. White paper.

Fishman, G. S., 1978. Principles of discrete event

simulation.

Varga, A., 2001. Discrete event simulation system. In Proc.

of the European Simulation Multiconference

(ESM'2001).

Howell, F. and McNab, R., 1998. SimJava: A discrete event

simulation library for java. Simulation Series, 30,

pp.51-56.

Jun, J.B., Jacobson, S.H. and Swisher, J.R., 1999.

Application of discrete-event simulation in health care

clinics: A survey. Journal of the operational research

society, pp.109-123.

Ng, K., Ghoting, A., Steinhubl, S.R., Stewart, W.F., Malin,

B. and Sun, J., 2014. PARAMO: a PARAllel predictive

MOdeling platform for healthcare analytic research

using electronic health records. Journal of biomedical

informatics, 48, pp.160-170.

Jamei, M., Nisnevich, A., Wetchler, E., Sudat, S. and Liu,

E., 2017. Predicting all-cause risk of 30-day hospital

readmission using artificial neural networks. PloS one,

12(7).

Bardhan, I., Oh, J. H., Zheng, Z. and Kirksey, K., 2014.

Predictive analytics for readmission of patients with

congestive heart failure. Information Systems Research,

26(1), pp.19-39.

Bradley, P. and Kaplan, J., 2010. Turning hospital data into

dollars: healthcare financial executives can use

predictive analytics to enhance their ability to capture

charges and identify underpayments. Healthcare

Financial Management, 64(2), pp.64-69.

Maguire, J. and Dhar, V., 2013. Comparative effectiveness

for oral anti-diabetic treatments among newly

diagnosed type 2 diabetics: data-driven predictive

analytics in healthcare. Health Systems, 2(2), pp.73-92.

Gartner, 2017. Gartner’s Top 10 Strategic Technology

Trends for 2017. http://www.gartner.com/smarterwith

gartner/gartners-top-10-technology-trends-2017/.

Färber, F., May, N., Lehner, W., Große, P., Müller, I.,

Rauhe, H. and Dees, J., 2012. The SAP HANA

Predicting Hospital Capacity and Efﬁciency

569

Database--An Architecture Overview. IEEE Data Eng.

Bull., 35(1), pp.28-33.

McGlothlin, J., Crawford, E., Srinivasan, H., Cianci, C.,

Bruneau, B., and Dorotta, I., 2017. Reducing Red Blood

Cell Transfusions. In Information Technology in Bio-

and Medical Informatics, pp. 120-133.

Smith, D. 2012. FDA: R OK for drug trials.Viewed October

2017. http://blog.revolutionanalytics.com/2012/06/fda-

r-ok.html

HEALTHINF 2018 - 11th International Conference on Health Informatics

570