A Bottom-up Simulation Method to Quantitatively Predict Integrated Care System Performance Contributed by: Zhengchun Liu, Dolores Rexachs, Francisco Epelde, Emilio Luque Presented by: Zhengchun Liu (
[email protected] or http://zliu.info ) At: 16th International Conference on Integrated Care, Barcelona High Performance Computing for Efficient Applications and Simulation Research Group (HPC4EAS)
Computer Architecture & Operating Systems Department
Universitat Autònoma de Barcelona
@
@
&
MOTIVATION
1
Prediction, explanation & optimization are challenging for a complex system like Integrated Care system. For example, healthcare operations management, for which we want to: Predict system performance for a specific configuration, cost and benefit for a proposed change. Explain factors influencing performance, how the prediction is made and why it performs like this. Optimize changes to the system with constrain like budget. 1. 2.
2 A platform to study healthcare system related problems, like bacteria propagation. (e.g., MRSA infection). To study disordered system behavior based on integration of first-principles model and datadriven model (real operation data). Every decision we make is based on information, stop guess.
The way to achieve to goal: First-principles modeling to capture details of system behavior from the interaction of system components.
MOTIVATION
1
Prediction, explanation & optimization are challenging for a complex system like Integrated Care system. For example, healthcare operations management, for which we want to: Predict system performance for a specific configuration, cost and benefit for a proposed change. Explain factors influencing performance, how the prediction is made and why it performs like this. Optimize changes to the system with constrain like budget. 1. 2.
2 A platform to study healthcare system related problems, like bacteria propagation. (e.g., MRSA infection). To study disordered system behavior based on integration of first-principles model and datadriven model (real operation data). Every decision we make is based on information, stop guess.
The way to achieve to goal: First-principles modeling to capture details of system behavior from the interaction of system components.
Start with simulating the emergency departments.
My Agenda ☛ Introduction ☛ The Emergency Department Simulator ☛ Use of the Simulator ☛ Model Parameters Calibration Tool ☛ Demo applications ☛ Conclusion and Future work
HOW
IT WORKS
Abbreviations:
Arrival
P Patient R Registration Staff Tr Triage Nurse DA Doctors in area A DB Doctors in area B NA Nurse in area A NB Nurse in area B A Auxiliary Staff I Medical Image L Laboratory Test LW BS Leave Without aaaaaaa Being Seen
I
By their own
P
Area B
R
Nursing P
Triage By Ambulance
P
Clinical Assessment
Waiting Room
Nursing P
A
> Length of Stay
Medical test P
DB
> Quality of Service
LWBS Further Test
AL 1,2,3
Area A
Treatment NA
Tr AL 4,5
I
Treatment
Yes Ambu. ?
NA
Pending
N
> ... > Root-cause > Sensitivity
Transferred
ED Medical Finished
Macro-Level Feature Admitted? Dead emergence
Cross Scenario Analysis
> ...
Home
Yes
"Sensoring"
P
L
DA
> Throughput
Sn ... ... S3 S2 S1
P
Laboratory Test
Understandable Knowledge of the Complex System
Exit
Clinical Assessment Medical test
Emergence
No
No
OR
STATE)
Registration
Medical Image Test P
(RULES + STATE VARIABLES =>
Conceptual Model
Hospital
Micro-Level Behavior Simulation Boarding
Further Test
Scenario Simulation
Note: Every patient who comes through the door is an unknown, with a condition that unfolds over time in a functionally nonInteracti deterministic way. Theoretically Informat speaking, no two paths through this “system” are the same for any two patients. sv1 p1 sv2 p2 st3 p3
Behavior
I/O
emulate the system behavior
Agent(Patient):
via
acuity level age body condition location …
I/O
Simulation Scenarios Interacting
State variables changed
S1 S2 S3 ... ... Sn
mimic individual’s behavior
sv2 p2
Variables: I/O
Interaction
sv1 p1
System Input
State transition
sv3 p3 ...
...
Parameters & State Variables
... ...
sv1 p1
Interacti on
State transition when interact with other agents or with time elapsing
sv1 p1 sv2 p2 st3 p3
sv2 p2
... ...
st3 p3 ... ...
sv1 p1 sv2 p2 st3 p3 ... ...
sv1 p1
sv1 p1
sv2 p2
sv2 p2
st3 p3
st3 p3
... ...
... ...
HOW
IT WORKS
Abbreviations:
Arrival
P Patient R Registration Staff Tr Triage Nurse DA Doctors in area A DB Doctors in area B NA Nurse in area A NB Nurse in area B A Auxiliary Staff I Medical Image L Laboratory Test LW BS Leave Without aaaaaaa Being Seen
I
By their own
P
Area B
R
Nursing P
Triage By Ambulance
P
Medical test P
Nursing
A
DB
LWBS Further Test
AL 1,2,3
P
First-Principle Models
Clinical Assessment
Waiting Room
Area A
Treatment NA
Tr AL 4,5
I
Treatment
Yes Ambu. ?
NA
Pending No
Clinical Assessment Medical test P
Laboratory Test N
Transferred
ED Medical Finished
Macro-Level Feature Admitted? Dead emergence
> ...
Home
Yes
"Sensoring"
P
L
DA
They are not as quick >and to build, but they Length ofeasy Stay have many advantages.> Quality In terms of Serviceof simulation, firstEmergence principle models provide extrapolation in addition Understandable > Throughput Knowledge of the to the interpolation > ...provided by data-driven Complex System models. >They Root-causealso can be used for prediction, > Sensitivity explanation and optimization. Cross Scenario Analysis Sn ... ... S3 S2 S1
Exit
No
OR
STATE)
Registration
Medical Image Test P
(RULES + STATE VARIABLES =>
Conceptual Model
Hospital
Micro-Level Behavior Simulation Boarding
Further Test
Scenario Simulation
Note: Every patient who comes through the door is an unknown, with a condition that unfolds over time in a functionally nonInteracti deterministic way. Theoretically Informat speaking, no two paths through this “system” are the same for any two patients. sv1 p1 sv2 p2 st3 p3
Behavior
I/O
emulate the system behavior
Agent(Patient):
via
acuity level age body condition location …
I/O
Simulation Scenarios Interacting
State variables changed
S1 S2 S3 ... ... Sn
mimic individual’s behavior
sv2 p2
Variables: I/O
Interaction
sv1 p1
System Input
State transition
sv3 p3 ...
...
Parameters & State Variables
... ...
sv1 p1
Interacti on
State transition when interact with other agents or with time elapsing
sv1 p1 sv2 p2 st3 p3
sv2 p2
... ...
st3 p3 ... ...
sv1 p1 sv2 p2 st3 p3 ... ...
sv1 p1
sv1 p1
sv2 p2
sv2 p2
st3 p3
st3 p3
... ...
... ...
configuration
HOW
IT WORKS?
-
SIMULATION INPUT ORGANIZATION
✓admission staff ✓triage nurse ✓nurse ✓doctor ✓auxiliary ✓carebox ✓laboratory test ✓internal test ✓external test ✓hospital ward ✓ambulance. ✓…
Two areas: A and B for different patients
Hospital ward
Version 2.0
patient(input)
scenario Scenario = ED-Model-Configuration + Input (Patient)
CTAS:
1
5
Patient:
CTAS: Canadian Triage and Acuity Scale
configuration
HOW
IT WORKS?
-
SIMULATION INPUT ORGANIZATION
✓admission staff ✓triage nurse ✓nurse ✓doctor ✓auxiliary ✓carebox ✓laboratory test ✓internal test ✓external test ✓hospital ward ✓ambulance. ✓…
Resource
Capacity (#) day night 3 2 2 0 3 1 2 1 2 4 5 5 2 5 4 4 5 2 4 2 50 60 3
Avg. Attention Time (AT, minutes) first interaction follow-up 5 3 8 6 20 15 15 13 25 18 20 14 8 7 6 5 11 7 7 5 45 30 15
Two areas: A and B for different patients
junior admission sta↵ senior admission sta↵ junior triage nurse senior triage nurse junior doctor in area A senior doctor in area A junior nurse in area A senior nurse in area A junior doctor in area B senior doctor in area B junior nurse in area B senior nurse in area B medical imaging test room laboratory test place carebox in area A chair in area B auxiliary nursing sta↵
Should Execute Many Times for One Scenario
AT Distribution Gamma Gamma Gamma Gamma exponential exponential exponential exponential exponential exponential exponential exponential Beta Beta exponential
Statistical Hospital ward Model
Version 2.0
patient(input)
scenario Scenario = ED-Model-Configuration + Input (Patient)
+
CTAS: Canadian Triage and Acuity Scale
HOW
IT WORKS?
-
SIMULATION OUTPUT CONFIGURATION
State information monitoring configuration
interaction information monitoring configuration
It is like: we could put a device (sensor) on each of the individuals to monitor their detailed activities. sensors are customizable and have process capability.
HOW
IT WORKS?
Extract
-
DIRECT SIMULATION DATA
Length of Stay, Occupancy, Length of Waiting, Efficiency, …
CALIBRATION -
AUTOMATIC TOOL
Purpose: Setting up a general model for the target system simulation; I.E., a general computational model TO specific ED simulator. Motivation: Enable the simulation users, e.g., ED manager, to calibrate parameters for their own ED system without the involvement of model developers. => promoting the application of simulation in ED studies. Challenge: Data Scarcity, Out the scope of Information System; Solution: Formed as an optimization problem; Process: selection of inputs, specifying the objective function, searching, and evaluating the calibration results
CALIBRATION -
such as length of stay), and part of the model parameters retrieved directly from real data. With respect to the unknown
SET UP YOUR OWN SIMULATOR (W HAT INFO . YOU NEED TO PROVIDE ) Although the empirical information is not accurate, it can dramatically reduce search space-size. Table 1 lists all the
parameters, empirical information such as boundary constraints, typical value can be obtained from experienced sta↵. parameters to be calibrated, as well as their boundary constraints. Thus the task is to search for an optimum set of
from your information system
from your experience
parameters which can lead to good (acceptable) fitness between the simulation results and actual data. Table 1: The parameters to be calibrated for the general agent-based model of emergency departments, in order to imitate the emergency department of Hospital of Sabadell . Note: LB and UB denotes Lower and Upper Boundary respectively, TV represents the Typical Value; all the units of time are in minutes. The Identity column corresponds to the circled numbers in Figure 1 denote the type of service.
Patient: arrival hour, day, acuity level, discharge time(date-time)
Identity Notation
System configuration: #doctor, #nurse, #labs (machine), #medical image, … (all about resource you have)
+
Description
LB
UB
TV
1
register T service
the parameter for registration service-time distribution model.
2
15
5
2
triage T service
the parameter for triage service-time distribution model.
5
20
10
3
nurseA T service
the average duration of service of nurses in area A.
8
30
16
4
doctorA T service
the average duration of service of doctors in area A.
8
30
18
5
nurseB T service
the average duration of service of nurses in area B.
5
20
12
6
doctorB T service
the average duration of service of doctors in area B.
5
20
15
7
imaging T service
the average duration for taking medical imaging.
20
40
25
8
lab T service
the average duration for taking laboratory test sample.
10
30
15
In summary, due to data scarcity, although the distribution of specific service duration cannot be fitted by such standard techniques as maximum likelihood estimation, we had some other time stamps which enable us to derive an
indirect approach to estimate the service-time distribution parameters. Our tool and general model
=
4. Model Calibration Calibration traditionally conceptualized as an step in model validation. It involves systematic adjustment of model parameters so that model outputs can accurately reflect the actual system behavior. To calibrate a model, three im-
value of parameters to set up your simulator (for your system) portant issues need to be addressed. The first issue is to select significant metrics to represent the emergent behavior
Example of uses, No. 1 The emergency department system is overcrowding, WHAT-IF we add 20 careboxes to the system?
Every decision we make is based on information, stop guess.
The influence of additional carebox on patients’ behavior (Area A).
benefits
the root-cause
Value
(a) length of stay
(b) length of waiting time (in treatment area)
Good?
(c) door-to-doctor time
Actual behavior
Macro (systemic level) Error
Our world is nonlinear
analytical prediction Singularity
Actual behavior
Extrapolation
Variables
Micro (component level)
Example of uses, No. 2
How can emergency departments respond to population aging: a simulation study.
1. Predict the effects of population aging on emergency department. 2. Make longterm plans and quantify their costs and benefits with the ED simulator. (explain) 3. Optimize changes to the ED system with constrain.
Information retrieved from real data (2014)
# tests
+
# Consultation
+ % H admission
=>
# LoS
Knowledge from actual data analysis: Elder patients need more care service and stay longer in ED.
Figure 5: Patient age distribution due to their severity.
Patients’ age distribution prediction model age year year NED (year) = Page · N · P · D re f age rate ED
Regarding that age PED
=
Replace Page ED in Equation 1 with Equation 2, we get:
age NED
Nre f ·
f Dre age
(1)
(2)
Zhengchun Liu et al. / for submitting for peer review xx (2016) 1–12 year Dage age age year NTable = reused · NinEDthis· Pmodel. rate ED (year) f 1: Notations Dage
8
(3)
With Equation 3, demography prediction and reference data in 2014, it is possible to roughly predict the number
Notations
Description
age NED
The number of patients due to age interval in 2014 (5 years in this study).
of patients will attend to the emergency department. The prediction will be given by patients age interval. With the demography information collected from INE, we made a ten-year patient arrival prediction from 2014 to 2023. The
Nreage Theoftotal number ofofpeople in the catchment area in ofFigure the hospital. f distribution arrival patients the prediction was illustrated 6.
*
Dyear distribution of5various age groups in the target area inincreased year (population pyramid). to see from Figure that, because of population aging,catchment the elder patient and younger patient age It is clearThe decreased. With emergency department simulator, and who the patient arrival prediction, it is possible to quantitively Page The probability of a person (due to age) will go to ED. ED evaluate year Prate
the response of the emergency department.
The ratio of population in year to the reference year (2014).
3. ED performance prediction Model assumption : The probability of a person
7
who will go to ED at least once per year depends on lots of factors, here we assume that the probability is depends on age and do not change over different year. That is to say, a fix probability will be used above the section studied of thepatient predictionattend of arrival aging in the future 15 years, and Figure 6 throughout the future toThe predict number topatients’ ED and, thecharacteristic age distribution of the patients.
*
illustrated the prediction. It is clear that the ED will have stress to deal with more old patients, this section will
Got from the Instituto Nacional de Estadística (INE) http://www.ine.es/
give the quantitive prediction of ED performance via simulation. Our previous work has developed a simulator
Patients’ age distribution prediction in the future
As input to the simulator to see QoS in the future
Knowledge from prediction: there will be more elder patients and less young patients attend ED
ED performance predicted by simulation Population aging
Patient arrival (input)
System configuration (resource)
scenario
Simulator and Analysis tool
System Performance (QoS)
Length-of-Stay
Leave-without-being-seen
Door-to-Doctor Time
Make plans in advance.
Population aging
proposal (changes)
Patient arrival (input)
System configuration (resource)
scenario
Propose longterm plans and quantify their cost and benefit with the ED simulator.
Simulator and Analysis tool
System Performance (QoS)
future performance without any changes
Effect of Changes?
future performance with: 3+ nurse in area A, 11+ careboxes
Door-to-Doctor Time
Length-of-Stay
future performance without any changes
future performance with: 3+ nurse in area A, carebox: 49 => 60
Leave without being seen
Optimize alternatives with constraint (e.g., budget, space) — work in process.
Conclusions & Future Work Conclusions: (1) A General Agent-Based Model for EDs (Spanish type); (2) Designed and Implemented an auto-calibration tool; (3) with this tool, we can have: Every decision we make is based on information, stop guess.
In summary, start from simulating the emergency departments, our efforts proved the feasibility and ideality of using agent-based model & simulation techniques to study healthcare system.
Future Work: (1) Population aging; How can emergency departments respond to population aging: a simulation study. (2) A step towards building a full model of integrated care system.
Thank you for Your Attention!
A Bottom-up Simulation Method to Quantitatively Predict Integrated Care System Performance Contributed by: Zhengchun Liu, Dolores Rexachs, Francisco Epelde, Emilio Luque Presented by: Zhengchun Liu (
[email protected] or http://zliu.info ) At: 16th International Conference on Integrated Care, Barcelona High Performance Computing for Efficient Applications and Simulation Research Group (HPC4EAS)
Computer Architecture & Operating Systems Department
Universitat Autònoma de Barcelona
@
@
&
