1. 1 Real-time Monitoring, Control and Optimization of Patients Flow in Emergency Departments Boaz Carmeli M.Sc. Thesis Exam Advisor: Prof. Avishai Mandelbaum The Faculty of Industrial Engineering and Management Technion - Israel Institute of Technology

2. 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. 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. 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. 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. 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. 7 Main Results – 1 (MIS) Monitoring and Measuring ED Load

8. 8 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. 9 Main Results – 2 (MIS and OR) Heuristic Solution for the ‘Which patient to treat next?’ problem

10. 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. 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. 12 Main Results – 3 (OR) Optimal analytic solution for the stylized fluid model of the PTN problem

13. 13 The Model

14. 14 Fluid Model Analysis We proved that a bang-bang control δ() defined over the set of time intervals T= {t s i, t e i } and over the time- varying arrival rate function α() as follows: δ(t) = µ, t s i <t<t e i, δ(t) = α(t-d) otherwise. Where ‘µ’ is the maximal service rate and ‘d’ is the deadline is optimal

15. 15 Fluid Model – Schematic View

16. 16 Constructing the Optimal δ()

17. 17 Main Results – 4 (OR) Applying realistic cost function for resolving the in-process decision of the PTN problem

18. 18 Cost Function Cost for triage scores: Triage 5Triage 4Triage 3 c 1 (t)=1*tc 1 (t)=2*tc 1 (t)=4*t Age distribution of patients: Over 7565-7545-65Under 45 C 2 (t)=5*c 1 (t)c 2 (t)=3*c 1 (t)c 2 (t)=2*c 1 (t)c 2 (t)= 1*c 1 (t) Expected ADT distribution: UnknownDischargedAdmitted c 3 (t)=1*c 2 (t) c 3 (t)= 2*c 2 (t)

19. 19 Cost Function - Continued Additional length of stay cost: If patient need to be admitted or ADT stats is unknown and is in the process more then 5 hours (300 minutes) If patient need to be discharged and is in the process more the 3.5 hours (210 minutes) c(t)= c 3 (t)+(t-300)^2c(t)= c 3 (t)+(t-210)^2

20. 20 Cost Function – Graphs Cost during the process (discharged only) Polynomial increase while getting to the maximal accepted LoS

21. 21 Thank You

22. 22 Real-time Monitoring, Control and Optimization of Patients Flow in Emergency Departments Boaz Carmeli M.Sc. Research Seminar Advisor: Prof. Avishai Mandelbaum The Faculty of Industrial Engineering and Management Technion - Israel Institute of Technology

23. 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. 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. 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 performance indicators (KPI) Analyze and interpret findings Control and Optimize Monitor and measure the environment Assess influence through monitor and measurement Adapt KPI to reflect new insights

26. 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) Ambulance Diversion Demand for ED Care Patient Arrive at ED Triage and room placement Diagnostic evaluation and ED treatment ED boarding of inpatients Leaves without treatment complete Patient disposition Ambulatory care system Transfer to other facility Admit to hospital Input: Arrivals Throughput: Under Treatment Output: Admitted/Discharged

27. 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 Load E. Vitkin, B. Carmeli, O. Greenshpan, D. Baras, Y. Marmor,

28. 28 System Architecture ED Simulator Based on observation Will be used, mainly, for design phase e.g. to mimic the RFID system RFID based Location Tracking 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 Learning Algorithms Analysis of Historical And Real-time Data Mathematical Models e.g. Queuing Theory Data Collection Analysis Data Visualization

29. 29 Real-time Monitoring Monitoring and Measuring ED Load

30. 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. 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

32. 32 Neural Networks Neural networks are learning networks calculating complex mathematical functions, which motivation came from the brain activity.

33. 33 Major advantage of NN approach is simplicity of different kinds: Easy to calculate Easy to present and understand (a tree-like hierarchy) Easy to modify Easy to adapt Neural Networks Advantages

34. 34 Network Structure We created network structure based on work of Solberg et al

35. 35 Input – Events Processing Network The Event Processing Network We use the EPN tool for collecting operational and clinical low level events

36. 36 Output - Dashboard

37. 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. 38 Load Tracking

39. 39 Real-time Control and Optimization Controlling and Optimizing the ED Patient Flow

40. 40 The ED Patient Flow Administrative Reception Triage & Vital Signs First Physician Examination Treatment Imaging (CT, MRI, US) ConsultingLab Tests Physician Decisions or Additional Tests Admit/Discharge/Transfer administration

41. 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. 42 Which Patient to Treat Next? (PTN) Patient under treatment at other ED ‘stations’ T5T5 T4T4 T3T3 T2T2 T1T1 Predicted Treatment P5P5 120 min5 P4P4 60 min4 P3P3 30 min3 P2P2 10 min2 P1P1 0 min1 Punishmen t Due Date TriageNew Arrivals Queues Content Doctor Triage Internal Queue Content Which patient should Doctor choose next?

43. 43 Queueing Model for the PTN Problem In Process patients Newly Arrived 2 In Process 2 physician Newly Arrived 1 Newly Arrived 3 In Process 1 In Process 3 In Process 4 Triage

44. 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. 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. 46 Performance Indicators Time Till First Encounter Meet triage deadline Total Length of Stay 4 hours

47. 47 Initial Analysis Serve next the patient with the longest waiting time. FCFS Serve next the patient with the highest waiting cost, based on some convex cost function, following the gcμ rule. Cost Chose from IP- patients using a cost function Strive to first meet NA-patient deadline constraint Hybrid

48. 48 Initial Results Total LoS Latent time Waiting time Service time Time till first encounter 176591031431 First come first serve 18959116148 NA patients first 178591051429 Dynamic Threshold

49. 49 Additional Justification Percentage of patients that meet the 4 hours LoS deadline Percentage of patients that meet the TTFE deadline 75%88% First come first serve 74%100% NA patients first 78%94% Dynamic threshold

50. 50 Parameters of a Typical ED Arrival Rates Encounter Distribution 65432Number of encounters 3%11%28%30%28%Percentage of patients Which means that 30% of the offered load is handled during first encounters

51. 51 Relevant Patient Parameters Triage scores: 543Triage Score 120 min60 min30 minDeadline 50%40%10%Distribution Age distribution of patients: Over 7565-7545-65Under 45 10%20%30%40% Expected ADT distribution: UnknownDischargedAdmitted 10%60%30%

52. 52 Cost Function Cost for triage scores: Triage 5Triage 4Triage 3 c 1 (t)=1*tc 1 (t)=2*tc 1 (t)=4*t Age distribution of patients: Over 7565-7545-65Under 45 C 2 (t)=5*c 1 (t)c 2 (t)=3*c 1 (t)c 2 (t)=2*c 1 (t)c 2 (t)= 1*c 1 (t) Expected ADT distribution: UnknownDischargedAdmitted c 3 (t)=1*c 2 (t) c 3 (t)= 2*c 2 (t)

53. 53 Cost Function - Continued Additional length of stay cost: If patient need to be admitted or ADT stats is unknown and is in the process more then 5 hours (300 minutes) If patient need to be discharged and is in the process more the 3.5 hours (210 minutes) c(t)= c 3 (t)+(t-300)^2c(t)= c 3 (t)+(t-210)^2

54. 54 Cost Function – Graphs Cost during the process (discharged only) Polynomial increase while getting to the maximal accepted LoS

55. 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. 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. 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. 58 Main Results – Dynamic Threshold Average length of stay for is 178 minuets

59. 59 Main Results – FCFS Average length of stay is 176 but no control on other indicators

60. 60 Results – Cost (age) The affect of age on the length of stay distribution throw the cost function

61. 61 Results – Cost (admitted/discharged) The affect of ADT on the length of stay distribution throw the cost function

62. 62 Time Till First Encounter Fix arrival rate Heavy traffic condition Triage 5 only Dynamic threshold algorithm

63. 63 Length of stay distribution

64. 64 Fluid Model Analysis We proved that a bang-bang control δ() defined over the set of time intervals T= {t s i, t e i } and over the arrival rate function α() as follow: δ(t) = µ t s i <t<t e i, δ(t) = α(t-d) otherwise Where ‘µ’ is the maximal service rate and ‘d’ is the deadline is optimal

65. 65 Fluid Model – Schematic View

66. 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. 67 Thank You

68. 68 More Backup Slides

69. 69 The ED Simulator We use the ED Simulator (developed by Dr. Marmor) for generating relevant input data into the system

70. 70 The Event Processing Network We use the EPN tool for collecting RFID data

71. 71 The Dashboard – Predicting ED Load

72. 72

73. 73 The Model Queues Two level decision Cost Constraints NA vs. IP Among IP Among NA NAIP

74. 74 The Monitoring and Control Dashboard – Example

75. 75 The ED Patient Flow