Models for Using Prediction to Facilitate Hospital Paitent Flow

MODELS FOR USING PREDICTION TO FACILITATE HOSPITAL PAITENT FLOW Jordan Peck1,4, James Benneyan3,4 Stephan Gaehde2, Deborah Nightingale1,4 1. Massachusetts Institute of Technology, 2. VHA Boston, 3. Northeastern University, 4. New England Veterans Engineering Resource Center Abstract The Emergency Department/Inpatient Unit Chain A common example of a health care delivery chain is the emergency department (ED) and inpatient unit (IU) chain. Recent studies have identified the output process of admitting a patient to a hospital IU as one of the greatest hindrances to flow in the ED [US GAO 2003, Olshaker & Rathlev 2006, Falvo et al. 2007, US GAO 2009]. A pull system (figure 1 a) is based on sharing information to closely match upstream production with downstream demand [Hopp & Spearman 2001]. It has been suggested that the ED/IU system can be improved by turning it into a pull system. In such a system, rather than wait for a bed to be ordered by the ED, when the IU opens a bed they would contact the ED and try to pull a patient. Although this may sound good in concept the ED does not have a stock of patients, and without knowing that patients are waiting in the ED, the IU staff does not have the incentive to open a bed and then pull a patient. By graphically representing the ED/IU chain in the same format (figure 1 b) it becomes clear that demand originates downstream instead of upstream. Therefore the downstream step, IU, would not know to pull a patient because they do not know the patient exists and instead they will focus primarily on the real demand of patients needing treatment. In response to this issue it has been suggested that if it can be predicted that a patient will need admission earlier in the ED treatment process, then the prediction can be shared with the IU which could begin the bed coordination process earlier and improve flow [Yen & Gorelick 2007]. This sharing of predictive information would create a quasi-pull system, where the order is passed downstream to mimic a supply pull system (figure 1 c). The concept above is contingent upon the ability to predict whether a patient is going to require admission to the ED. A recent study has developed a method of assigning a probability of admission to each patient that enters the ED. At any moment the probabilities of all patients in the ED can be added to create running expected bed need index [Peck et al. 2011]. The value of sharing this index with the IU can then be tested. Recent studies of emergency department (ED) patient flow have identified the process of moving patients from an ED into a hospital inpatient unit (IU) as the primary bottleneck in the hospital health care delivery chain. Some of these studies have suggested predicting whether a patient will require hospital admission when they enter the ED. These predictions allow for advanced planning and enable a pull system. However, given the large amount of variability in the hospital system the flow benefits of using a predictive method relative to existing systems (ED utilization warning and discharge by noon heuristics) is unknown. This paper discusses the concept of a pull system in the hospital and uses simulation to study how it may benefit emergency department flow. Introduction A health care delivery chain can be summarized in many cases as: Admit, Treatment Step 1, Treatment Step 2, through treatment step N and Discharge. It is the connection of steps/processes through which a patient flows in order to receive their treatment, and enables a new dimension of analysis for optimizing health care delivery. The comparison of health care delivery chains to supply chains suggests certain methods for improving flow and performance: forecasting, queuing analysis, facility design, process improvement, scheduling, and simulation modeling. Many of these methods have indeed been used for improving flow in a health care system and are generally discussed in the field of “patient flow” [Hall 2006]. However, when considering a chain of events, the connectedness of steps means that flow based decisions need not be set before the patient enters the system. Instead these decisions are dependent on the states of other steps within the chain. For example it has been shown that the quality of supply stocking policies often improves when the state of the system is continuously reviewed [Simchi-Levi et al. 2003, Cachon & Terwiesch 2009]. The increase in review periods often leads to pull systems. Consequently, many who study patient flow suggest that pull systems are needed in health care [Graban 2008]. 1 Figure 1 a) Traditional Pull Sys ystem, b) Current ED/IU Push System, c) Potential ED/IU “Pull” “Pu System the West Roxbury campus (VHA (VH Boston). The logic of the model is shown in figure ure 2 and is comprised of four primary sub-models: Arrival, Ar the Emergency Department, the Inpatient Unit it and Bed Management. The arrival sub-modell consists c of a creation module that generates patient ents based on the actual patient arrival pattern derived ed from the VHA Boston data. After a patient is create ated they are assigned a probability of being admitted.. This T was done using the data from the prediction study y that t has been performed at VHA Boston [Peck et al. 2011]. 201 This data was used to create a distribution of pro robabilities of admission for assigning a probability to each simulated patient that enters the ED. Each of the he patient probabilities are then summed to provide thee imperfect i predicted bed need index. Patients then move mo through a decision module which assigns whether er the patient will indeed Methods To some, the discussion of creati ating a prediction based pull system in the ED may ay have intuitive benefits. However since human livess aare at risk when changes are made in health systems,, simulation is a popular tool for exploring “what if” f” scenarios [Hall 2006]. One type of simulation that is com ommonly used for studying the ED/IU system iss discrete event simulation (DES) [Baesler et al. 200 003, Connelly & Bair 2004, Kolb et al. 2008, Jingshann & Howard 2010, Paul et al. 2010, Peck & Kim, 2010] 10]. The model in this study was built in Rockwell Aut utomation, Inc.’s ARENA DES software version 13.5 .5. The model is based on the ED/IU chain at the V Veteran’s Health Administration Boston Health Care Sy System, located in 2 require admission, which is used to ge generate a perfect predicted bed need index. Both of the he bed indexes are shared with the bed man anagement sub model. Figu gure 2 Discrete Event Simulation Model Logic Upon receiving their admissio ssion predictions, patients enter the ED which is compri prised of 13 beds, just like the VHA Boston ED. Patien tients then seize a bed for a treatment duration based on the distribution of treatment times observed at the actu ctual VHA Boston ED. While the simulation is run unning, ED bed utilization is being shared with bedd m management. At this point those patients who we were chosen for admission enter a queue to seize an IU bed while continuing to hold an ED bed andd their admission prediction is updated to a 1. Disc ischarged patient admission predictions are reduced to 0. The running admission indexes are updated accordin rdingly. When a patient enters the IU, they hey are assigned a length of stay based on the data col collected from the VHA Boston hospital. To capture hhow information can effect decisions and consequently ly flow, the model assumes that the doctor is the decis cision maker and limited resource. This is modeled as in figure 3. As can be seen in the figure a patient first irst seizes a doctor for treatment. The patient can only nly be treated or discharged by this unique doctor from rom that point on. The patient then goes through som some randomized amount of value added treatment that at ddoes not require the doctor. At the end of the cycle the he amount of time the patient spent is deducted from the patient’s total value added IU LOS. The patient will ill continue to go through this cycle until they havee ccompleted their treatment. At this point, rather tha than re-enter the treatment queue, the patient enters the he queue to seize a doctor for discharge orders. In this wa way those patients waiting for discharge are in directt competition for doctors with patients who are still rece ceiving treatment. However, to manage this compet petition, the bed management module has the abilityy tto shift priority between the two processes when certai tain conditions are met. No o Doctor Treat Patient Treatment Compl Complete? Yes Doctor Discharge Patient Value Added Intermediate Treatment or Tests Figure 3 Doctor Decision De Cycle Simulation Scenarios Four primary scenarios were w studied using this simulation. These scenarios are ar facilitated using the bed control sub-model which h sets se priority to discharge based on: 1. 2. 3. 4. me of day priority is set to Time: At a predetermined time discharge for three hours. ED crowding: While a designa nated number of ED beds have been occupied. Imperfect Prediction: While a designated, imperfectly predicted, bed index level is reached. rea Perfect Prediction: While a designated, perfectly predicted, bed index level is reached. rea Each of these scenarioss can c then be tested for sensitivity using factors thatt have h been built into the model. In this case, sensitivity ity to a non-value added admission delay between the ED and IU was tested. Results ults n for two cases: no non All results below are shown value added admission delay ay and a 30 minute non value added delay. The erro rror bars in each figure 3 Figure 6 shows how ED waiting time changes as sensitivity to the perfectly predicted bed need index decreases. represent the 95% confidence interval for the data point based on 100 replications of the simulation. Figure 4 shows how ED wait time changes with the time of day that discharges are emphasized. Each discharge emphasis shift last for 3 hours. ED Wait Time: No NVA Admission Delay 0.8 ED Wait time: No NVA Admission Delay 0.6 0.8 0.4 0.6 0.4 0.2 0.2 0 1 0 18 17 16 15 14 13 12 11 10 9 8 7 0.6 0.4 0.4 0.2 0.2 0 1 0 9 8 7 5 6 7 8 9 10 11 12 13 2 3 4 5 6 7 8 9 10 11 12 13 Figure 6 ED Wait Time by Perfect Bed Need Index Sensitivity Figure 4 ED Wait Time with Shifting Discharge Emphasis Time Figure 7 shows how ED waiting time changes as sensitivity to the imperfectly predicted bed need index decreases. Figure 5 shows how boarding time changes as sensitivity to occupied beds decreases. ED Wait Time: No NVA Admission Delay 0.8 4 0.6 0.8 18 17 16 15 14 13 12 11 10 3 ED Wait Time: 30 min NVA Admission Delay 0.8 ED Wait time: 30 min Admission Delay 2 ED Wait Time: No NVA Admission Delay 0.8 0.6 0.6 0.4 0.4 0.2 0.2 0 0 1 2 3 4 5 6 7 8 1 9 10 11 12 13 3 4 5 6 7 8 9 10 11 12 13 ED Wait Time: 30 min Admission Delay 0.8 ED Wait Time: 30 min Admission Delay 0.8 2 0.6 0.6 0.4 0.4 0.2 0.2 0 0 1 2 3 4 5 6 7 8 1 9 10 11 12 13 2 3 4 5 6 7 8 9 10 11 12 13 Figure 7 ED Wait Time by Imperfect Bed Need Index Sensitivity Figure 5 ED Wait Time with ED Bed Utilization Sensitivity 4 emphasize discharge and times when it is not. Therefore by only activating a sensitivity level for a specific hour will show varying results depending on the hour. Figure 9 shows how the system performs when the sensitivity is set to 5 for one specific hour of the day. As can be seen there are some hours that cause the system to perform well and some that do not. Discussion From the results above, it is clear that the simulation has a great deal of variability. Although this makes analysis of data more difficult, this was planned to make the system more realistic. Beginning with the results in figure 4 it can be seen that the waiting time does decrease as the emphasis of discharge is moved earlier with some notable exceptions. As would be expected from the cyclical nature of the simulation and of a true hospital there are key points of resonance that will either dampen or exacerbate waiting time. This may be the result of doctors missing patient treatment which causes them not to be seen until the next day (afternoon peak). It is also possible that a doctor misses seeing the patient early enough such that the patient cannot be seen as many times in a day (early morning peak). The resonant peaks can be seen in all other results as certain sensitivities drive the system to operate in that zone. With that said, beside the resonant peaks, the ED crowding and the bed need index scenarios can all be adjusted such that there is a significant improvement in flow. Figure 8 shows the best performers of each scenario, where discharging between 8 and 11 can be called discharge by noon to fit with the common flow improvement method [ACEP 2008]. As can be seen by the figure, perfect index performs the best. Also, in the case where there is a 30 minute non value added discharge delay, the crowding and imperfect information scenarios are able to compensate for the delay and eliminate its effects while discharge by noon cannot. 0.1 0.09 0.08 0.07 0.06 0.05 0.04 0.03 0.02 0.01 0 1.2 1 0.8 0.6 0.4 0.2 0 ED Wait Time: Admit Index Sensitivity… 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 Figure 9 Time Varied Effect of Imperfect Index Sensitivity = 5 Each sensitivity level displays a different effectiveness pattern. Therefore it is possible to vary the index sensitivity by time of day such that the best performing sensitivity is active for each hour. Figure 10 shows the optimal sensitivity variation for the imperfect information index for the system with no delay. Using this pattern resulted in an average waiting time less than the best perfect index scenario of constant level as shown in figure 8. Prioritize Discharge ED Waiting Time No NVA Admission Delay 30 Min Discharge Delay Prioritize Treatment By Noon ED Imperfect Perfect Crowding Index Index Optimal Indexing Figure 10 Optimal shifting imperfect bed need index sensitivity for simulated system. Figure 8 Best Performing Points of Each Scenario Conclusions The results show that sensitivity to the prediction indexes can be adjusted such that there is a net improvement in flow. However based on the time scenario there are times of day when it is good to Emergency Department flow is a long studied issue that remains a chief concern to many hospitals. 5 Investigating Emergency Department Overcrowding. SIMULATION, 86(8-9), pp.559 -571. Peck, J.S. & Kim, S.-G., 2010. Improving patient flow through axiomatic design of hospital emergency departments. CIRP Journal of Manufacturing Science and Technology, 2(4), pp.255-260. Peck, J. et al., 2011. Predicting emergency department inpatient admissions to improve same-day patient flow. Submitted for Publication. Rossetti, M.D., Trzcinski, G.F. & Syverud, S.A., 1999. Emergency department simulation and determination of optimal attending physician staffing schedules. In Simulation Conference Proceedings, 1999 Winter. Simulation Conference Proceedings, 1999 Winter. IEEE, pp. 1532-1540 vol.2. Simchi-Levi, D., Kaminsky, P. & Simchi-Levi, E., 2003. Designing and managing the supply chain: concepts, strategies, and case studies, McGraw Hill Professional. US GAO, 2003. Hospital Emergency Departments: Crowded Conditions Vary among Hospitals and Communities. GAO-03-460. US GAO, 2009. Hospital Emergency Departments: Crowding Continues to Occur, and Some Patients Wait Longer than Recommended Time Frames. GAO-09-347. Yen, K. & Gorelick, M.H., 2007. Strategies to Improve Flow in the Pediatric Emergency Department. Pediatric Emergency Care, 23(10), pp.745-749. Recent studies have suggested that prediction can be used to drive behavior in the inpatient unit and create a pull system to improve flow. This paper showed that, in a simulated hospital, prediction does indeed have the ability to improve flow and reduce the effects of non-value added delays. References ACEP et al., 2008. Emergency Department Crowding: High Impact Solutions. no. American College of Emergency Physicians. Baesler, F.F., Jahnsen, H.E. & DaCosta, M., 2003. Emergency departments I: the use of simulation and design of experiments for estimating maximum capacity in an emergency room. In Proceedings of the 35th conference on Winter simulation: driving innovation. WSC ’03. Winter Simulation Conference, pp. 1903–1906. Connelly, L.G. & Bair, A.E., 2004. Discrete Event Simulation of Emergency Department Activity: A Platform for System‐level Operations Research. Academic Emergency Medicine, 11(11), pp.11771185. Falvo, T. et al., 2007. The Opportunity Loss of Boarding Admitted Patients in the Emergency Department. Academic Emergency Medicine, 14(4), pp.332-337. Graban, M., 2008. Lean hospitals: Improving quality, patient safety, and employee satisfaction, Productivity Press. Hall, R.W., 2006. Patient flow: reducing delay in healthcare delivery, Springer. Hopp, W. & Spearman, M., 2001. Factory physics: foundations of manufacturing management, McGraw-Hill/Irwin. Jingshan Li & Howard, P.K., 2010. Modeling and analysis of hospital emergency department: An analytical framework and problem formulation. In 2010 IEEE Conference on Automation Science and Engineering (CASE). 2010 IEEE Conference on Automation Science and Engineering (CASE). IEEE, pp. 897-902. Kolb, E.M.W. et al., 2008. Reducing emergency department overcrowding: five patient buffer concepts in comparison. In Proceedings of the 40th Conference on Winter Simulation. Miami, Florida: Winter Simulation Conference, pp. 1516-1525. Olshaker, J. & Rathlev, N., 2006. Emergency department overcrowding and ambulance diversion: the impact and potential solutions of extended boarding of admitted patients in the emergency department. Journal of Emergency Medicine, 30(3), pp.351-356. Paul, S.A., Reddy, M.C. & DeFlitch, C.J., 2010. A Systematic Review of Simulation Studies Biographies Jordan Peck is a PhD candidate in MIT’s Engineering Systems Division. He is a research assistant in MIT’s Lean Advancement Initiative and holds and MS in Technology and Policy. James Benneyan, PhD, is the director of Northeastern University’s Health Care Systems Engineering Program and a leading authority in the field. Stephan Gaehde, MD, is the Chief of the VA Boston Emergency Medicine Service and holds an MPH with a focus in Medical Informatics Deborah Nightingale, PhD, is a Professor of Practice in Aeronautics and Astronautics Department and Engineering Systems at MIT. She is the director of MIT’s Lean Advancement Initiative and Center for Technology, Policy and Industrial Development. 6