1
Real-time Monitoring, Control and Optimization of Patients Flow in
Emergency Departments
Boaz CarmeliM.Sc. Thesis Exam
Advisor: Prof. Avishai MandelbaumThe Faculty of Industrial Engineering and ManagementTechnion - Israel Institute of Technology
2
Thesis Focus Areas
Improving ED operations, focusing on real-time aspects: Real-time monitoring Real-time load measurement Real-time patient-flow related
decisions
3
Thesis Structure
Understand ED operations End-to-end Strategic, operational, real-time
Identify key performance indicators Design a prototype ED real-time
monitoring and control system Based on event processing paradigm
Real-time and adaptive load monitoring
4
Thesis Structure - Continued
Understand real-time patient-flow within the ED Which patient should physician treat next?
Follows ED manager insights and directives
Simulate ED patient flow Simulated data - the ‘faculty’ ED simulator Compare various patient-flow scheduling
algorithms Analyze a fluid model for the ‘which patient
to treat next’ problem Propose an optimal solution
5
Main Results
Propose an original approach for ED load monitoring
Together with Edward Vitkin Formulate the ‘which patient to treat next?’
(PTN) problem Cost vs. constraint approach Cost functions
Propose multiple solutions for the PTN problem Heuristic solution
Proactive computing Analytic solution for stylized fluid-model
time-varying arrival rates Analytic solution for steady-state heavy traffic
Contributed insights and problem understanding
6
Publications
Vitkin E., Carmeli B., Greenshpan O., Baras D., Marmor Y., MEDAL: Measuring of Emergency Departments Adaptive Load, MEDINFO 2010.
Huang, J., Carmeli, B., Mandelbaum, A., 2012. Control of Patient Flow in Emergency Departments, or Multiclass Queues with Deadlines and Feedback. In Preparation.
Carmeli B., Mandelbaum A., Promise: Real-time Patient Flow Monitoring and Control in Emergency Departments, The IEM 2010 Conference
7
Main Results – 1 (MIS)
Monitoring and Measuring ED Load
8
Monitoring and Measuring ED Load
We defined a framework which provides a mean to monitor and measure load
The framework is based on Neural Networks paradigm which enables adaptive load definition
A NN learning mechanism adapts the load function towards specific ED setting and user (e.g., patient, physician) views
9
Main Results – 2 (MIS and OR)
Heuristic Solution for the ‘Which patient to treat next?’ problem
10
Dynamic Control
At point of decision: Check if any of the triage patients are just
about to miss their deadline If so – server triage patients
Else – perform a look ahead into the triage queues to check if all waiting patients can be served before their deadlines:
Assume you will serve the triage queues with all available capacity till all of them will be served
If look ahead check succeed Serve the IP-patients
Otherwise Serve the triage patients
11
Dynamic Control - Continued
If triage patients was chosen Choose the one that is closest to the
deadline (waiting-deadline) If already beyond the deadline chose
patient with highest fraction waiting/deadline
If in-process patients was chosen Chose the one with the highest
waiting cost Using the provided cost functions
12
Main Results – 3 (OR)
Optimal analytic solution for the stylized fluid model of the PTN problem
13
The Model
14
Fluid Model Analysis
We proved that a bang-bang control δ(•) defined over the set of time intervals T= {ts
i, tei} and over
the time-varying arrival rate function α(•) as follows:
δ(t) = µ, tsi<t<te
i ,
δ(t) = α(t-d) otherwise.Where ‘µ’ is the maximal service rate and ‘d’ is the deadlineis optimal
15
Fluid Model – Schematic View
16
Constructing the Optimal δ(•)
17
Main Results – 4 (OR)
Applying realistic cost function for resolving the in-process decision of the PTN problem
18
Cost Function
Cost for triage scores:
Triage 3Triage 4Triage 5
c1(t)=4*tc1(t)=2*tc1(t)=1*t
Age distribution of patients:
Under 4545-6565-75Over 75
c2(t)= 1*c1(t)
c2(t)=2*c1(t)
c2(t)=3*c1(t) C2(t)=5*c1(t) Expected ADT distribution:
AdmittedDischargedUnknown
c3(t)= 2*c2(t) c3(t)=1*c2(t)
c3(t)=1*c2(t)
19
Cost Function - Continued
Additional length of stay cost:
If patient need to be discharged and is in the process more the 3.5 hours (210 minutes)
If patient need to be admitted or ADT stats is unknown and is in the process more then 5 hours (300 minutes)
c(t)= c3(t)+(t-210)^2 c(t)= c3(t)+(t-300)^2
20
Cost Function – GraphsCost for Admitted patients
0
500
1000
1500
2000
2500
3000
3500
4000
4500
5000
0 50 100 150 200 250
Time (min)
Co
st
t3 a1 d1
t3 a2 d1
t3 a3 d1
t3 a4 d1
t4 a1 d1
t4 a2 d1
t4 a3 d1
t4 a4 d1
t5 a1 d1
t5 a2 d1
t5 a3 d1
t5 a4 d1
-20000
0
20000
40000
60000
80000
100000
120000
0 100 200 300 400 500 600
t3 a1 d1
t3 a1 d2
t3 a2 d1
t3 a2 d2
t3 a3 d1
t3 a3 d2
t3 a4 d1
t3 a4 d2
t4 a1 d1
t4 a1 d2
t4 a2 d1
t4 a2 d2
t4 a3 d1
t4 a3 d2
t4 a4 d1
t4 a4 d2
t5 a1 d1
t5 a1 d2
t5 a2 d1
t5 a2 d2
t5 a3 d1
t5 a3 d2
t5 a4 d1
t5 a4 d2
Cost during the process (discharged only)
Polynomial increase while getting to the maximal accepted LoS
21
Thank You
22
Real-time Monitoring, Control and Optimization of Patients Flow in
Emergency Departments
Boaz CarmeliM.Sc. Research Seminar
Advisor: Prof. Avishai MandelbaumThe Faculty of Industrial Engineering and ManagementTechnion - Israel Institute of Technology
23
What you are about to see
The problem – monitoring and control of ED operations
Core IT concepts of ED monitoring and control system
Two interesting applications: ED load monitoring and measurement
Just a very brief overview due to time constraints
Which patient to treat next? Core ED patient flow control challenge
24
The Problem
The rising cost of healthcare services has been a subject of mounting importance and much discussion worldwide
Overcrowding in hospital Emergency Departments (ED) is perhaps the most urgent operational problem in the healthcare industry
Overcrowding in hospital EDs leads to excessive waiting times and repellent environments, which in turn cause:
Poor service quality (clinical, operational) Unnecessary pain and anxiety for patients Negative emotions (in patients and escorts) that sometimes lead
to violence against staff Increased risk of clinical deterioration Ambulance diversion Patients leaving without being seen (LWBS) Inflated staff workload
25
Solution Approach
Improve ED operation efficiency through real-time monitoring and control
taking clinical, operational and service level aspects into account (Sample) Key Performance Indicators:
Time Till First Encounter Total Length of Stay Bed Occupancy ED Load And many more…
Define the key Define the key performance performance
indicators (KPI)indicators (KPI)
Analyze and Analyze and interpret interpret findingsfindings
Control Control and and
OptimizeOptimize
Monitor andMonitor andmeasure themeasure theenvironmentenvironment
Assess influence through monitor and measurement
Adapt KPI to reflect new insights
26
ED Conceptual Model
Emergency Care•Seriously ill and injured
patients from the community
•Referral of patients with emergency conditions from other providers
Unscheduled urgent care
•Desire for immediate care
•Lack of capacity for unscheduled care in the ambulatory care system
Safety net care•Vulnerable populations (eg,
Medicaid beneficiaries, the uninsured) care
•Access barriers (eg, financial, transportation, insurance, lack of usual source of care)
AmbulanceDiversion
Demand forED Care
Patient Arriveat ED
Triage and room
placement
Diagnostic evaluation
and ED treatmentED
boarding of
inpatients
Leaves without
treatment
complete
Patientdispositio
n
Ambulatory
caresystem
Transfer to other facility
Admit to hospital
Input: Arrivals
Throughput: Under Treatment
Output: Admitted/Discharged
27
Real-time ED Monitoring and Control System
Data Collection Collect real-time relevant information from hospital IT
systems such as PACS, EHR, ADT, LAB etc Adding RFID based location tracking system for Physicians,
Nurses, Patients and other relevant personnel Data Visualization
Operational dashboard Displays complex behaviors in a simple way
Mobile devices Analysis Techniques
Mathematical models – service engineering Simulations – for planning and control Machine learning - neural networks, based on historical data
Published paper (MedInfo 2010): MEDAL: Measuring of Emergency Departments Adaptive LoadE. Vitkin, B. Carmeli, O. Greenshpan, D. Baras, Y. Marmor,
28
System Architecture
ED Simulator•Based on observation
•Will be used, mainly, for design phase e.g. to mimic the RFID system
RFID based LocationTracking
•Low level location tracking for patients and care personnel
•Technology dependent capabilities
Hospital IT systems•Admit, Discharge, Transfer
•Electronic Health Records
•Lab request/results
•Picture Archive and Communication System (PACS)
Real Time Event Processing
Network Rule Based
Analysis
Machine LearningAlgorithmsAnalysis of Historical
And Real-time Data
Mathematical Models
e.g. Queuing Theory
Data Collection Analysis Data Visualization
29
Real-time Monitoring
Monitoring and Measuring ED Load
30
ED Load
What is ED Load? Number of people in ED? Number of waiting people in ED? Percent of time Doctor/Nurse works? Combination of these?
Clearly, there is no one simple answer.However, there are more questions, which can guide us
toward the desired result:
What are the factors affecting Load? How can we combine them? Do we have to have same Load Definition for different EDs? Do we have to have same Load Definition for different
duties?
31
Monitoring and Measuring ED Load
We defined a framework which provide a mean to monitor and measure load
The framework is based on Neural Networks paradigm which enables adaptive load definition
A NN learning mechanism adapts the load function towards specific ED setting and user (e.g., patient, physician) views
36
Output - Dashboard
37
Learning User Needs
Since user feeling of the system is not an explicit function we provide him tool for “easy” feedback:
INCREASE increase decrease DECREASE
38
Load Tracking
39
Real-time Control and Optimization
Controlling and Optimizing the ED Patient Flow
40
The ED Patient Flow
Administrative Reception
Triage & Vital Signs
First Physician Examination
Treatment Imaging(CT, MRI, US)
Consulting Lab Tests
Physician Decisions or Additional Tests
Admit/Discharge/Transfer administration
41
Controlling the ED Patient Flow
Modeling the ED patient flow as a queueing network Patients – tasks Care personal – servers (stations)
Knowing in real-time the next ‘station(s)’ in the patient’s route Set of alternatives are usually provided by the care personnel
No a priory full path knowledge System may provide decision support
Deciding upon the ‘best’ next station (e.g. next physician) Assuming there are multiple options Sends patient to the (clinically and operationally) ‘best’ station Always make sure there is at least one ‘next’ station
Within each ‘station’ queue deciding upon the next patient to treat
Based on operational, clinical and patient fairness service level aspects
42
Which Patient to Treat Next? (PTN)
Patient under treatment at other
ED ‘stations’
T5
T4
T3
T2
T1
Predicted
Treatment
P5120 min5
P460 min4
P330 min3
P210 min2
P10 min1
Punishment
Due Date
Triage
New Arrivals Queues Content
Doctor
Triage
Internal Queue Content
Which patient should Doctor choose next?
43
Queueing Model for the PTN Problem
In Process patients
Newly Arrived 2
InProcess 2
physician
Newly Arrived 1
Newly Arrived 3
InProcess 1
InProcess 3
InProcess 4
Triage
44
The ED Manager View
Reduce total length of stay at the ED while meeting triage deadlines Try to keep total length of stay below 4
hours for all patients Is this an appropriate goal?
Take service level aspect into account Patient’s age
Precedence to old patients Expected discharged/admitted aspect
Precedence to patients that are expected to be discharged to their home after treatment
45
Addressing the Clinical View
We suggest a service policy that seek to reduce the overall waiting cost while meeting triage deadlines Minimal effort due-date policy Allocate as much efforts as possible
for IP-patients, following the known generalized cμ rule
46
Performance Indicators
Time Till First Encounter Meet triage deadline
Total Length of Stay 4 hours
47
Initial Analysis
FCFSServe next the patient with the longest waiting time.
CostServe next the patient with the highest waiting cost, based on some convex cost function, following the gcμ rule.
Hybrid
Strive to first meet NA-patient deadline constraint
Chose from IP-patients using a cost function
48
Initial Results
Time till first encounter
Service time
Waiting time
Latent time
Total LoS
First come first serve
311410359176
NA patients first
81411659189
Dynamic Threshold
291410559178
49
Additional Justification
Percentage of patients that meet the TTFE deadline
Percentage of patients that meet the 4 hours LoS deadline
First come first serve
88%75%
NA patients first
100%74%
Dynamic threshold
94%78%
50
Parameters of a Typical ED
Arrival Rates Encounter
Distribution
Number of encounters23456
Percentage of patients 28%
30%
28%11%3%
Which means that 30% of the offered load is handled during first encounters
51
Relevant Patient Parameters
Triage scores:
Triage Score
345
Deadline30 min60 min120 min
Distribution
10%40%50% Age distribution of
patients:Under 4545-6565-75Over 75
40%30%20%10%
Expected ADT distribution:
Admitted
Discharged
Unknown
30%60%10%
52
Cost Function
Cost for triage scores:
Triage 3Triage 4Triage 5
c1(t)=4*tc1(t)=2*tc1(t)=1*t
Age distribution of patients:
Under 4545-6565-75Over 75
c2(t)= 1*c1(t)
c2(t)=2*c1(t)
c2(t)=3*c1(t) C2(t)=5*c1(t) Expected ADT distribution:
AdmittedDischargedUnknown
c3(t)= 2*c2(t) c3(t)=1*c2(t)
c3(t)=1*c2(t)
53
Cost Function - Continued
Additional length of stay cost:
If patient need to be discharged and is in the process more the 3.5 hours (210 minutes)
If patient need to be admitted or ADT stats is unknown and is in the process more then 5 hours (300 minutes)
c(t)= c3(t)+(t-210)^2 c(t)= c3(t)+(t-300)^2
54
Cost Function – GraphsCost for Admitted patients
0
500
1000
1500
2000
2500
3000
3500
4000
4500
5000
0 50 100 150 200 250
Time (min)
Co
st
t3 a1 d1
t3 a2 d1
t3 a3 d1
t3 a4 d1
t4 a1 d1
t4 a2 d1
t4 a3 d1
t4 a4 d1
t5 a1 d1
t5 a2 d1
t5 a3 d1
t5 a4 d1
-20000
0
20000
40000
60000
80000
100000
120000
0 100 200 300 400 500 600
t3 a1 d1
t3 a1 d2
t3 a2 d1
t3 a2 d2
t3 a3 d1
t3 a3 d2
t3 a4 d1
t3 a4 d2
t4 a1 d1
t4 a1 d2
t4 a2 d1
t4 a2 d2
t4 a3 d1
t4 a3 d2
t4 a4 d1
t4 a4 d2
t5 a1 d1
t5 a1 d2
t5 a2 d1
t5 a2 d2
t5 a3 d1
t5 a3 d2
t5 a4 d1
t5 a4 d2
Cost during the process (discharged only)
Polynomial increase while getting to the maximal accepted LoS
55
Dynamic Control – Informal Description
At point of decision: Check if any of the triage patients are just
about to miss their deadline If so – server triage patients
Else – perform a look ahead into the triage queues to check if all waiting patients can be served before their deadlines:
Assume you will serve the triage queues with all available capacity till all of them will drained out
If look ahead check succeed Serve the IP-patients
Otherwise Serve the triage patients
56
Dynamic Control - Continue
If triage patients was chosen Choose the one that is most close to
the deadline (waiting-deadline) If already beyond the deadline chose
patient with highest portion waiting/deadline
If In-process patients was chosen Chose the one with the highest
waiting cost E.g., apply gcμ rule
57
Main Observations
In most situations there are enough physicians at the ED to serve triage patients exactly at their deadline
Look ahead provides additional proactive action towards extreme arrival rates
There may be situations in which triage patients will still miss there deadline
58
Main Results – Dynamic Threshold
Average length of stay for is 178 minuets
Time Indicators by Triage Groups
135
169
195
178
55
94
125
105
12.523.7
37.229.1
14.0 15.0 15.4 15.1
67 60 57 59
0
50
100
150
200
250
Triage - 3 Triage - 4 Triage - 5 All
Tim
e (m
in)
Avg LoF
Avg Waiting
Avg TTFE
Avg Service
Avg Latent
Meeting the deadlines
14.3
20.1
26.0
22.4
11.9
5.2 5.5 5.9
0.0
5.0
10.0
15.0
20.0
25.0
30.0
Triage - 3 Triage - 4 Triage - 5 All
Triage categories
% fr
om a
ll pa
tient
s
% > 4 Hours
% Missed deadline
59
Main Results – FCFS
Time Indicators by Triage Grouls
153162
190
176
8492
114103
24.3 28.334.1 31.0
16.4 15.1 14.9 15.1
52 5663 59
0
20
40
60
80
100
120
140
160
180
200
Triage - 3 Triage - 4 Triage - 5 All
Tim
e (m
in)
Avg LoF
Avg Waiting
Avg TTFE
Avg Service
Avg Latent
Meeting the Deadlines
25.0
18.9
28.2
24.225.0
17.8
5.1
11.9
0.0
5.0
10.0
15.0
20.0
25.0
30.0
Triage - 3 Triage - 4 Triage - 5 All
% o
f p
atie
nt'
s ca
teg
ory
% GT 4 Hours
Missed the deadline
Average length of stay is 176 but no control on other indicators
60
Results – Cost (age)
Time Indicators by Age Groups
200
174162
154
178
131
96 97
67
105
27.1 30.0 29.4 32.8 29.115.4 15.9 13.9 14.1 15.1
5563
53
7659
0
50
100
150
200
250
Under 45 45 - 65 65 - 75 Over 75 All
Tim
e (
min
)
Avg LoF
Avg Waiting
Avg TTFE
Avg Service
Avg Dormant
The affect of age on the length of stay distribution throw the cost function
Meeting the deadlines(Age groups)
29.0
20.8
16.7 17.0
20.9
6.84.2
6.37.5
5.9
0.0
5.0
10.0
15.0
20.0
25.0
30.0
35.0
Under 45 45 - 65 65 - 75 Over 75 All
% > 4 Hours
% Missed deadline
61
Results – Cost (admitted/discharged)
The affect of ADT on the length of stay distribution throw the cost function
Time Indicators by ADT Groups
185171
197
178
119
95
123
105
28.7 29.1 30.6 29.114.3 15.4 15.6 15.1
5562 59 59
0
50
100
150
200
250
Admitted Discharged Unknown All
Tim
e (
min
)
Avg LoF
Avg Waiting
Avg TTFE
Avg Service
Avg Latent
Meetinf the deadlines(ADT groups)
22.3 22.124.5
23.0
6.5 5.7 5.7 5.9
0.0
5.0
10.0
15.0
20.0
25.0
30.0
Admitted Discharged Unknow n All
Tim
e (m
in) % > 4 Hours
% Missed deadline
62
Time Till First Encounter Fix arrival rate Heavy traffic condition Triage 5 only Dynamic threshold algorithm
Triage 5 - TimeTill First Encounter
0
10
20
30
40
50
60
70
80
0 10 20 30 40 50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200
Time (min)
Nu
mb
er o
f ar
riva
ls
63
Length of stay distribution
Length of stay
0
5
10
15
20
25
30
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
Length of Stay (Hours)
Len
gth
of
stay
(%
)
LoS
64
Fluid Model Analysis
We proved that a bang-bang control δ(•) defined over the set of time intervals T= {ts
i, tei} and over the
arrival rate function α(•) as follow: δ(t) = µ ts
i<t<tei,
δ(t) = α(t-d) otherwiseWhere ‘µ’ is the maximal service rate and ‘d’ is the deadlineis optimal
65
Fluid Model – Schematic View
66
Summary
Advances in information technology and usability call for better utilization of computer based monitoring and control systems within hospitals and specifically within the ED
Rambam recently extended their EHR into the ED
Digital data collection and monitoring open the door for utilizing traditional as well as newly developed operations research and service science methodologies to be used within hospitals
We identified several potential points for improving the ED operations, researched and analyzed two of them:
Adaptive load monitoring and measurement
Dynamic control for improving patient flow i.e., by answering the question: which patient should physician treat next?
67
Thank You
68
More Backup Slides
69
The ED SimulatorWe use the ED Simulator (developed by Dr. Marmor) for generating relevant input data into the system
70
The Event Processing NetworkWe use the EPN tool for collecting RFID data
71
The Dashboard – Predicting ED Load
72
73
The Model
Queues Two level decision
Cost Constraints
NAvs.IP
AmongIP
AmongNA
NA IP
74
The Monitoring and Control Dashboard – Example
75
The ED Patient Flow