SIMULATION-BASED MODELS OF EMERGENCY DEPARTMENTS: REAL-TIME CONTROL, OPERATIONS PLANNING AND SCENARIO ANALYSIS

Transcription

1 SIMULATION-BASED MODELS OF EMERGENCY DEPARTMENTS: REAL-TIME CONTROL, OPERATIONS PLANNING AND SCENARIO ANALYSIS Sergey Zeltyn 2, Yariv Marmor 1, Avishai Mandelbaum 1, Boaz Carmeli 2, Ohad Greenshpan 2, Yossi Mesika 2, Segev Wasserkrug 2, Pnina Vortman 2, Dagan Schwartz 3, Kobi Moskovitch 3, Sara Tzafrir 3, Fuad Basis 3, Avraham Shtub 1, Tirza Lauterman 1 1 Technion Israel Institute of Technology, 2 IBM Haifa Research Labs, 3 Rambam Health Care Center ABSTRACT The Emergency Department (ED) of a modern hospital is a highly complex system that gives rise to numerous managerial challenges, spanning the full spectrum of operational, clinical and financial perspectives. Since realistic ED models are intractable analytically, one resorts to simulation for an appropriate framework to address these challenges, which is what we do here. Specifically, we apply a general and flexible ED simulator to address several central wide-scope problems that arose in a large Israeli hospital. First, we demonstrate that our simulation model can support realtime control by inferring missing data of the current ED state, which then enables short-term prediction and operational planning (e.g. nurse staffing). To this end, we implement a novel simulation-based technique that utilizes the concept of offered-load. Then, using the same simulationbased approach, we evaluate the impact of RFID (Radio Frequency Identification) technology on ED operational metrics and costs. Finally, we analyze design and staffing problems that arose from physical relocation of the ED, which lead to the implementation of design and process improvements. A prerequisite for all of the above is an extensive (cleaned and validated) hospital data-based system, which is the data source for our simulations, presently offline and potentially (after implementing an RFID system) in real-time. 1 INTRODUCTION 1.1 Operations management in Emergency Departments: Main challenges and simulationbased modeling The rising cost of healthcare services has been a subject of mounting importance and much discussion, worldwide. Ample reasons have been proposed, for example increasing life spans and the availability of an ever-increasing number of costly diagnostic and therapeutic modalities (Hall et al. 2006). Yet, regardless of their cause, rising costs impose, and rightly so, pressures on healthcare providers to improve the management of quality, efficiency and the economics in their organizations. A critical healthcare organization, widely recognized in need of urgent enhancements, is the large hospital, the complexity of which is well represented by the micro-cosmos of its Emergency Department (ED). The latter is our focus here for being the window through which a hospital is judged for better or worse, and for amplifying a variety of problems that arise also elsewhere. ED management should intertwine clinical, operational and financial dimensions. In this paper, 1

2 we focus on a somewhat operationally-biased (business process) view, which is then expanded to accommodate interactions with the other clinical and financial aspects. From an operational view, overcrowding and subsequent excessive delays are the most urging ED problems (Sinreich and Marmor 2005), having clear interactions also with ED clinical and financial dimensions. Citing Green, 2008, arguably, the most critical delays for healthcare are the ones associated with healthcare emergencies. Overcrowding in the ED can and does cause, among other things, the following (see, for example, Derlet and Richards, 2000): Poor service (clinical) quality: Patients with a severe problem (e.g. undiagnosed myocardial infraction) can wait for hours until their first diagnosis by a physician (which could become life threatening). Other patients are getting treatment that is inferior to the one they would have gotten after being properly diagnosed and hospitalized in the appropriate wards. Patient in unnecessary pain: When ED staff is too busy, patients are often neglected to experience unnecessary pain or discomfort - there could simply be no one able to approach them, for example, when the staff is catering to more clinically-urgent cases. Negative emotions, all the way to violence against staff: Extended waiting times, combined with an overcrowded environment and psychological pressures, is a recipe for agitation and violent behavior. Ambulance diversion: Over-congested EDs could turn incapable of accepting newly arriving ambulances, which gives rise to ambulance diversion and its multitude negative ripple effects. Patients' LWBS (Leave Without Being Seen): Some patients, being exhausted by waiting, abandon the ED at different phases of their process (often to be returning in later times and worsened conditions). Inflating staff workload: The longer the ED sojourn the longer the ED effort required (for example, as is the case in our partner hospital, when the protocol calls for a nurse-visit every 15-minutes of a patient's ED stay). Increased vulnerability: Long sojourns increase the likelihood of clinical deterioration, contagion of additional maladies and, all in all, the occurrence of adverse events. There exist tools and methods that help to alleviate the problem of overcrowding and excessive waits. These solutions typically demand careful planning of the ED processes and appropriate staffing scheduling techniques for nurses and physicians. See Sections and for a discussion on related work and Section that introduces a new approach to staff scheduling. However, it turns out that, in order to apply these methods properly, one should first resolve several other critical problems. Some of these problems are studied in our paper and are briefly specified below. Availability of information on the current ED state. Proper functioning of the ED, even given that adequate workforce and technical resources are available requires precise information on the current state of the ED. Specifically, we have in mind at least the accessibility to reliable information on the number and profile of patients of different types in the ED, the patients' state in the ED process (e.g. if results of a certain laboratory test are available for a specific patient) and accessibility of data on the physical location of patients (e.g., when a patients in a bed returns from an X-ray and placed in a location that differs from the one she was originally taken from). At least partially, this data should be available via the hospital IT systems. However, the data in these systems frequently turns out unreliable (Ash, Berg and Coiera, 2004) as it is fed by humans, who tend to circumvent or ignore procedures and thus fail to provide updates in real time. (We hasten to add that, in the hospital setting, such data-maintenance 2

3 failures are often the outcome of clinical emergencies taking their well-deserved priorities.) Moreover, some types of data cannot be extracted from most of the currently installed IT system. For example, assume that a patient is waiting in a queue for an X-Ray check: what is the patient's place in queue and the anticipated wait until service start? In our paper, we shall address the challenges that arise from such unreliable and inaccessible data by using an online simulation approach (Section 4.2), and examining the benefits and costs of RFID (Radio Frequency Identification) technology (Section 5), which enables patients location tracking in realtime. Short-term forecasting and operations planning. As a rule, operational decisions, including staff scheduling, are preferably planned in advance. In this paper, we consider the problem of short-term planning over a future horizon of several hours-to-days. Two problems should be solved in order to enable effective planning. First, one should implement an adequate forecasting model that predicts the number of exogenous arrivals to the ED. Such a model should provide predictions for each patient type since different types conceivably require different treatments and amounts of work from ED resources. Second, one should develop a model that combines the forecasts of external arrivals with the internal dynamics of the ED. Such a model would support operational decision making throughout the ED. Providing short-term predictions within the context of a command and control system presents its own unique challenges, in particular analysis must be performed within a short time frame (in the order of minutes), and must be based on data regarding the current ED state. Section 4 below is dedicated to this issue. Scenario analysis and strategic planning. Assume that design changes are planned for the ED. Two concrete examples, considered in this paper, are physical re-location of the ED (Section 6) and evaluation of an RFID-based system for patients location tracking (Section 5). Given such proposed design changes, ED management should evaluate how the ED processes will be altered, what will be the consequences for staff scheduling and what are the financial costs to be incurred. Such considerations must be taken into account when deciding on whether to implement contemplated design changes or, in case the changes are avoidable, how to implement them in the best possible way. Simulation-based modeling of the ED. All challenges formulated above require a model of the ED. As analytical models are unable to capture the complexity of ED operations, a major component of our solution is an ED simulation model (as reported in Sinreich and Marmor 2005 and discussed in Section 3). It turns out that our simulation-based model is general and flexible enough to address these challenges. It provides estimates regarding the current operational state, completing the missing data; it can incorporate forecasting and staffing techniques that enable high-quality short term operational planning. Moreover, it can be smoothly integrated with the real-time decision support system that we are currently developing (Greenshpan et al. 2009). Finally, it is very useful in scenario analysis towards supporting strategic planning. 1.2 Contribution and structure of the paper In subsequent sections, we continue with the survey of related work (Section 1.3) and describe the ED of an Israeli hospital where our models are applied (Section 2). We now proceed with an outline of our research contributions. 3

4 1.2.1 General simulation model of the ED In this paper, a simulation-based modeling approach is applied to a wide spectrum of ED problems. It is important to emphasize that all these problems were addressed using a single general flexible simulation model of the ED (Sinreich and Marmor 2005, 2006). Slight modifications of the base model were required for each specific problem but its flexible design enables quick and efficient adaptation, on demand. Significantly, although all our case studies were performed in a single Israeli hospital, the model can be easily modified and applied to other EDs. (In fact, the work of Sinreich and Marmor was based on research performed in the EDs of eight hospitals, leading to the creation of a "universal" ED simulator.) Section 3 of our paper contains a description of our simulation model Short-term forecasting and operations planning In Section 4, we apply our simulation-based approach to real-time control, short-term forecasting and operations planning. Starting with a brief problem statement in Section 4.1, we continue to simulation-based inference of the current state in Section 4.2. We emphasize especially the problem of inferring patients' discharge times. (At discharge times, patients are either released home or transferred to hospital wards.) As it turns out, hospital information on these discharge times is not precise and we complete it via simulation. Section 4.3 describes forecasting of external arrivals to the ED. Then, in Section 4.4, our two main staff scheduling methods are presented. We start with the prevalent RCCP (Rough Cut Capacity Planning) method (Vollman, Berry and Whybark 1993) and then present a new refined technique, based on the concept of offeredload. The offered load of a resource type (nurses, physicians, imaging devices), at a given time, is its (average) amount of work in process, where work is measured in time-units. This concept refines RCCP in the sense that it allocates workload accurately over time (while RCCP, on the other hand, accounts for all the workload brought in by a patient right at the arrival time of that patient). The offered-load concept originates in queueing-theory (see Feldman et al. 2008, and references therein); it is here adapted to the complex ED environment. Roughly speaking, the application of our offered load approach involves two stages. First, a simulation with an "infinite" number of resources is run; this yields, for each resource type, estimates of its time-dependent offered load (which assumes out delays due to scarce resource). The offered-load gives rise to a nominal time-dependent staffing level, required from each resource type. It provides the baseline for calculating actual staffing levels, accounting for performance goals and resource constraints. (Feldman et al develops successful staffing strategies, based on the offered-load, but merely to single-queue systems.) Section 4.6 presents results of our simulation experiments. We start with validation of our current state inference approach, comparing simulation-based estimates with real data from the ED database. Satisfactory useful results are reported. Then, a simulation-based forecasting model is run and load estimates for physicians and nurses during the future 8 hours are derived, using RCCP and the offered load approaches. Finally, we assume that recommendations that follow the RCCP and offered load approaches were in fact implemented. We then run new simulations with the corresponding staffing levels of physicians and nurses. The results of these simulationsindicate that the offered load approach is preferable over the RCCP approach. Remark. Content of Section 4 is based on the conference paper of Marmor et al. (2009). 4

5 1.2.3 Scenario analysis and strategic planning: costs and benefits of RFID technology In Section 5, we proceed to scenario analysis, estimating possible benefits of RFID implementation in the ED. We start with the description of two candidate RFID technologies in Section 5.1. Both technologies and an ideal technology that combines their advantages are used in our simulation modeling. In Section 5.2, the required process changes are outlined. Specifically, our goals are to prevent unauthorized customers departures (LWBS Left Without Being Seen), eliminate or decrease unnecessary waiting times of customers that are forgotten in Imaging (CT and X-Ray) units and, finally, expose possible problems in the physical layout of the ED. Results of comprehensive simulation experiments, presented in Section 5.3, reveal that RFID implementation under the assumptions of our model affects the Average Length of Stay (ALOS) in two opposite directions. Decrease of unnecessary waiting times, naturally, implies ALOS decrease. However, decrease or elimination of LWBS customers increases the number of patients in the system, hence also workload and ALOS. Our simulation experiments help decision-makers to estimate and compare both of these effects. Concerning operational performance of the two RFID technologies we have found that Passive RFID technology provides a very good answer to our problems. However, the ultimate decision on RFID technology implementation must be left to the hospital decision-makers, who could take into account additional considerations: implementation cost, layout constraints, psychological factors related to the use of different RFID systems etc Integration with decision-support system in ED In order to provide ED managers and other decision-makers the appropriate tools for real-time control and operational planning, one should design an interface between our simulation models and the IT systems of the hospital. In addition, ED data should be presented to the decisionmakers in an efficient and convenient way. Currently, we are developing a Decision Support System that addresses this challenge (Greenshpan et al. 2009). In Section 4.7, we demonstrate how this system can be integrated with real-time control and forecasting features, as presented in Section 4. If RFID technology is introduced into the ED, the Decision Support System can be significantly enhanced - Section 5.4 elaborates on this issue Scenario analysis and strategic planning: ED re-design under physical relocation A decision to design and construct a new ED was taken by management of our partner Israeli hospital. As the first stage in that transition, the ED was moved to a temporary location at the basement of the hospital, allowing the old ED to be renovated and expanded. Then, in 2010, a new permanent ED will be opened at the location of the previous ED. Our simulation-based approach has been used to evaluate the consequences of these two ED transfers, from different perspectives: staff scheduling, staff walking distances, design and performance of imaging facilities etc. In Section 6, we present the main research issues of a project that was dedicated to the transfer from the original location to the temporary location. Many of our recommendations, that arose from this research, were actually implemented. (A project on the second stage of transfer, to the permanent location, was also carried out but is not reported here.) Our main conclusions are as follows: 5

6 Staffing analysis of the nursing team was performed. (See Section 6.2.) It turned out that the nurses treating Internal Walking patients would become a bottleneck unless process changes are performed. In the initially planned state, this nurse team had to treat all walking patient (Internal, Surgical and Orthopedic) located in the same ED ambulatory unit. We recommended to schedule an additional nurse to the Ambulatory unit during most hours of the day, and relocate a nurse from the Trauma (Surgical & Orthopedic) unit to the internal unit. The ED, both in its previous and temporary state, has its own X-Ray unit. Patients are either sent to this unit or routed to the general X-ray unit of the hospital. We modeled several scenarios for ED X-Ray operations, under alternative operating hours and patients routing policies (see Section 6.3 for a discussion). Consequently, the initial suggestion to open the ED X-Ray room over the extended period of 10 or even 24 hours of the day was rejected. The optimal opening hours turned out to be 12:00-18:00. During hours when the ED X-Ray unit is closed, all patients are sent to the general X-ray unit. Concerning the patients routing policy during 12:00-18:00, it was recommended to send patients to the general X-ray unit according to a threshold strategy, specifically once the length of queue to the ED X-ray exceeds seven patients. Since the area of the temporary ED is significantly larger that the area of the previous ED (2,000 square meters versus 1,000 respectively), it turned out that the walking distances of the staff increase, on average. Consequently, slightly higher staffing levels should be used in the temporary ED, in order to sustain the same service level, as before. In addition, we recommended some changes in the ED design: related ED units with large flow of staff and patients between them should be located as close as possible to each other. For example, adjacent location of Trauma patients room and Treatment room have a high priority and it is preferable to move the Ambulatory ED, where Walking patients reside, closer to Trauma ED. (Some physicians treat both types of patients; therefore, their walking distances could be reduced.) Simulation-based comparison between the originally planned state of the ED and the state after implementation of our recommendations (specified above and some others) has shown that the full implementation of our recommendation implies improvement of all significant ED performance measures. For example, Average Length of Stay (ALOS) of some patient types decreased by nearly 60 minutes. Sensitivity analysis with respect to arrival rates have shown that the temporary ED, working under our recommendations, can function under 10% load increase and still provide better service level than under its originally planned design. 1.3 Related work Simulation in support of ED operations The application of simulation has been instrumental in addressing the multi-faceted challenges that the healthcare domain is presenting (Kuljis, Paul, and Stergioulas 2007). Wide spectrum of ED problems also received significant attention in this kind of research. It is quite common to use simulation, mostly by researchers, to compare operational models or to assess a model that addresses a specific research question. For example, Medeiros, Swenson, and DeFlitch 2008, present a simulation-based validation of a novel approach to ED processes change, placing an emergency care physician at triage. Kolb et al. 2008, study different policies of patient transfer from ED to Internal Wards, in order to decrease the resulting over- 6

7 crowding and delays. (Tseytlin 2009, solves a similar problem for our Israeli hospital using analytical approaches, based on queueing models.) For some reviews on simulation-based approach in health care, see Jun, Jacobson, and Swisher (1999), White (2005) and Jacobson, Hall and Swisher (2006). Improvement of patients experience in EDs via application of simulation and Lean Manufacturing tools was considered in Khurma, Bacioiu, and Pasek (2008). The prevalent approach for addressing ED overcrowding is staff (re)scheduling (e.g. Sinreich and Jabali (2007), Badri and Hollingsworth (1993), namely adding or shifting in time staff resources so as to uniformly maintain acceptable ED performance (e.g. time to the First Encounter with a Doctor, or FED time). Most such works focus on off-line steady-state decision making, as opposed to on-line operational and tactical control. Other researchers analyze alternative operational ED designs (Garcia et al. 1995; King, Ben-Tuvim, and Bassham 2006; Liyanage and Gale 1995) for example, comparing acuteness-driven models (e.g. triage) against operations-driven models (e.g. fast-track, which assigns high priority to patients with low resource requirements). A widespread approach is to decompose the problem by focusing only on one type of resource. An example is an effort to schedule nurses while ignoring the scheduling of other resources (Draeger, 1992); or scheduling physicians and nurses, one after the other (Sinreich and Jabali, 2007). These attempts, based on simulation models, predict performances of the ED as a function of staffing and scheduling decisions. The simulation models require input in the usual form of patient arrivals and service durations, of each patient by each resource type, exactly as in the simulation that we are using. We are, however, unaware of any uses of simulation in a hospital setting for real time command and control. Nor are we aware of any work in which simulation has been used to complete partial data regarding the current operational state. These research directions are pursued in Section 4. In a broader perspective, our research gives rise to a multitude of practical and theoretical challenges, many of which touch on active simulation-driven research. For example, input modeling (Biller and Nelson 2002) and historical (trace-driven, resampling) simulation (Asmussen and Glynn 2007; McNeil, Frey and Embrecht 2005), are both related to the problem of properly incorporating actual ED data into our simulator. Yet, deserving of an expanded attention is Symbiotic simulation (Fujimoto et al. 2002, Huang et al. 2006), defined as "one that interacts with the physical system in a mutually beneficial way", "driven by real time data collected from a physical system under control and needs to meet the real-time requirements of the physical system" (Huang et al. 2006). Additionally (Fujimoto et al. 2002), symbiotic simulation is "highly adaptive, in that the simulation system not only performs "what-if" experiments that are used to control the physical system, but also accepts and responds to data from the physical system". In some of our ED implementations, however, the interaction between the simulator and its underlying physical system must go beyond the common symbiotic simulation framework (see Section 4). Specifically, we obtain real-time data regarding current state, then complete the data when necessary via simulation, next predict short-term evolution and workload, and finally proceed with simulation and mathematical models as decision support tools, all this in real-time or close to real-time. 7

8 1.3.2 Alleviating overcrowding: analytical approaches to staff scheduling Although simulation-based approach is in the focus of our research, we emphasize that an optimal approach to real-life ED problems should combine simulation and analytical insights. These insights can be especially valuable when staff scheduling problems must be solved. In general, both deterministic and stochastic mathematical methods can be applied. For example, Beaulieu et al. (2000) present a deterministic mathematical programming approach to staff scheduling. The RCCP approach, demonstrated in Section 4 (Vollmann, Berry, and Whybark 1993), is also based on deterministic considerations. However, in our opinion, stochastic models, based on queueing theory, are more appropriate for capturing volatile and inherently non-deterministic ED reality. Although, it is hard to design a tractable comprehensive queueing model for ED, it is possible to apply simpler models combining them with simulation. The research on the offered load concept, presented in Section 4 provides us with an example of this approach. Using the technique, applied to time-varying queueing systems in Feldman et al. (2008), we develop the staff scheduling algorithm that jointly uses simulation and analytical staffing formulae. See Green (2008) for further references on these and related issues Applications of RFID technology in health care Significant research and development efforts have been devoted to the search after efficient and accurate Indoor Location Tracking (ILT) systems. While the Global Positioning System (GPS) has become the de-facto standard for outdoor tracking, and it serves as the foundation for many location tracking applications, GPS has yet no equivalent leading technology, which is suitable for indoor tracking (Lee et al. 2006). ILT systems are also referred to as RFID, after the technology of Radio Frequency IDentification. RFID technology has recently become widespread due to its many merits. Basically, RFID provides unique identifications to objects, hence it can be used as the foundation for objects tracking, monitoring and control (Hightower and Borriello 2001; Hightower, Want and Borriello 2000). RFID has traditionally been used for tracking passive entities, such as consumer package goods, medications and medical equipment. Yet this same technology can be used for uniquely identifying humans, e.g. patients and care personnel in hospitals. Applying RFID for indoor location tracking requires an additional layer, which associates the RFID tag with a specific location. This association can be implemented via two conceptually different approaches (Saha et al. 2003): Cell-based location tracking location identified through the location of the reader of the RFID tag. Triangulation location calculated from radio frequencies, used in the communication between the RFID tag and scattered RFID readers (Bahl and Padmanabhan 2000). RFID-based ILT systems have been recently developed for addressing specific needs that arise in patients' care. For example, MASCAL (Emory and Leslie 2005) is an integrated solution for tracking patients and equipment during events of mass causality; MASCAL is based on the communication network, and it is integrated with the hospital's clinical database. As another example, an RFID-based system was deployed in Taiwan (Wang et al. 2006), for identification and tracking of potential SARS cases; the system provides active patient-location tracking information as well as body temperature indication. In this present work, RFID it the technology 8

9 behind our proposed ILT systems enabling data-based business process management - in particular transformation towards improvement. 2 RESEARCH FRAMEWORK This research is a part of an Open Collaborative Research program, a combined research effort of three organizations partnered together: the Faculty of Industrial Engineering & Management at the Technion Institute, IBM s Haifa Research Laboratory and a government-affiliated Israeli hospital which is Israel s largest northern medical center, catering to over 2 million citizens (about one third of Israel s population). The hospital comprises 36 Internal Wards, around 1,000 patients can be hospitalized simultaneously and 75,000 patients are hospitalized yearly. In this research project, we focus on several hospital units including the ED which is the gate and the window to the hospital, and which must operate in a mass-customized mode i.e., follow a structured care process while providing to each individual the specific care required. The ED of our partner Israeli hospital accepts 82,000 patients per year, with 58% classified as Internal patients and 42% as Surgical or Orthopedic patients. Mean sojourn time of patients in the ED equals 4:38 hours, with a large variance over individual patients. The simulation-based approach turned out to be well-tuned for operational and strategic challenges that arise in the ED. Our recommendations, that arose from the application of this approach, were successfully applied when the ED was moved from the permanent to a temporary location (Section 6). Other research issues that are studied in Sections 4 and 5: real-time control, decision support system for operational planning and validation of RFID technology implementation, provide promising research results and will, hopefully, be followed by full-scale implementation. 3 BASIC SIMULATION MODEL OF THE EMERGENCY DEPARTMENT In Figure 1, we depict two perspectives of the care process that patients undergo at the ED: the resource (i.e. physicians, nurses, etc.) perspective, and the process (activities) perspective. In this care process, two types of queues portray the delays that patients experience: first are resource queues (rectangular, in red), which are due to limited resources (e.g. nurses, imaging equipment); the second are synchronization queues (triangular, in green), which arise when one process activity awaits another (e.g. a patient waiting for results of blood tests and x-ray, in order to proceed with the doctor's examination). The care process in an ED was captured in a simulation model, created with the generic simulation tool of Sinreich and Marmor (2005). In addition to the care process, the simulation model requires patients arrival processes, for each patient type, and staffing levels of the medical staff, with their respective skills. For our purpose, the model was configured to the ED specs of our partner hospital, as follows. There are six types of patients, which also require different skills from the caring Physicians. Patient types 1 and 2, which are Internal Acute and Internal Walking respectively, are treated by internal physicians. Patient types 3 and 4, which are Surgical Acute and Surgical Walking respectively, require treatment by surgical physicians. Finally, Patient types 5 and 6, Orthopedic Acute and Orthopedic Walking respectively, require an orthopedic physician. Acute patients need a bed while walking patients use chairs. In addition, patient types differ by the arrival process (e.g. number of arrivals per hour and by day-of-week), and by the decisions made in the patient care process (e.g. the percentage of patients sent to X-ray). 9

10 Figure 1 ED resource-process chart The actual simulation tool is comprised of the following three modules: 1. The first module is a Graphical User Interface (GUI) that describes the general unified process illustrated in Figure 2. Through the GUI, the user can input data and customize the general process to fit the specific ED modeled and receive operational results from the ED after simulation run. 2. The second module includes two mathematical models that are used to estimate patient arrivals and staff walking time. The simulation tool uses the models for patient arrival estimation that were developed in Sinreich and Marmor (2005). 3. The third and final module is the simulation model itself. This model receives data from both the GUI and the mathematical models. The simulation is updated and customized automatically to fit a specific ED based on data and information the user passes on to the GUI. The simulation model is transparent to the user who is only required to interact with a user friendly GUI without the need to learn a simulation language syntax. 4 SIMULATION-BASED MODELING FOR REAL-TIME CONTROL AND OPERATIONS PLANNING IN ED 4.1 Research goal In this section, we start to apply our simulation-based modeling approach to real-life ED problems. We show that this approach can help to ED managers infer the missing information on the current ED state, provide a reliable forecast of the ED state in the short-term and perform opera- 10

11 tional staff scheduling decision. Finally, in Section 4.7 we demonstrate how our simulationbased tools are integrated with the Decision Support Systems for hospital managers. 4.2 Simulation-based inference of current state As discussed in Section 1.1, reliable information on the current state of ED is crucial for the realtime control and operational planning. Typically, only partial data of the current ED state is maintained and available from the hospital's electronic data systems. For example, in our case, no data exists regarding the queue (number) of patients waiting to be seen by a physician. One expects the amount and quality of usable data to constantly improve over time, due to the introduction of additional data entry systems or new technologies (e.g. sensor technologies, such as RFID and ultra-sound, for accurate location tracking of patients, staff and equipment). However, within the chaotic ED environment, it is reasonable to expect that some data will always remain unavailable or too costly to acquire. We now discuss how to infer missing data, using the simulation model described above. Such simulation-based inference must deal with several issues. The first is consistency: how to generate simulation paths that are consistent with available ED data. Another important issue is data inaccuracy. (Note that inaccurate data adds complexity to generating simulation realizations that are consistent with the provided data.) A third challenge, arising due to the availability of only incomplete data, is the identification of an appropriate initial state for the simulation. The way we overcome this last hurdle is to feed in actual arrival data for a long enough period of time that ensures that the simulation warm-up period is over, prior to estimating the missing data. Coping with consistency and inaccuracy raises interesting research questions, as already alluded to. Here we content ourselves with two ED-specific practical examples, of accommodating actual ED data accurate and inaccurate. Accurate data - taking actual arrivals into account: In our partner ED, receptionists enter data into the IT systems, in particular regarding patient arrivals, as part of the admittance process. The medical state of the majority of arriving patients is such that they actively participate in the registration process, as the first step upon arrival. Registration of the others, acute patients unable of self-registration, is carried out by the paramedics bringing them in, shortly after arrival. Therefore, arrivals data accurately captures actual patients' arrival times it can be thus fed as is into the simulator. (Receptionists also record patient type - Internal, Surgical, or Orthopedic - upon arrival.) To this end, we modified, in obvious manners, our generic simulator, which originally generates arrivals as a stochastic process (Poisson or relatives). It can now generate realizations consistent with the arrival data, when the latter is fed from an external database. Inaccurate data - taking discharges into account: Data about patients' discharge (departure) time, in our partner hospital, may be inaccurate. Specifically, each departure time is registered by the receptionist upon completion of the ED treatments the patient is then ready to leave, for either home or to other hospital wards. In the (common) case when there is no ward immediately available to accept the patient, inaccurate data arises. Then, patients spend additional time waiting in the ED, which not only goes unrecorded but it also influences subsequent beds/chairs occupancy and ED staff utilization (due to time spent on catering to these delayed patients). Additional inaccuracies occur due to patients leaving without being seen (Green 2008), with or without their medical files, and some other accounting-related reasons. We found no efficient way for generating simulation realizations that are consistent with our discharge data, except for discarding inconsistent simulation paths. Note, however, that the probability of generating a realization in which the simulated departure times correspond exactly to 11

12 the provided departure times is negligible. To this end, and to overcome both inaccuracy issues, we condition on the number of patients of each type that were discharged from the ED according to the data. Namely, we considered a (short-term) simulation realization to be consistent if, at the end of the simulation run, the number of patients that were allowed discharge (of each type) equals, to within some accuracy constant, the number of patients of this type that were discharged according to the data. The results turned out satisfactory though, clearly, more thought is required here. In Section 5, we shall perform scenario analysis for the RFID technology implementation and further explore the issues related to benefits of exact knowledge of the ED state. 4.3 Forecasting ED arrivals For simulating an ED future evolution, one must simulate patients arrivals to the ED. Figure 2, from our partner hospital, demonstrates that ED arrival rates strongly depend on day-of-week and hour-of-day. In addition, holidays and days after holidays have unusual patterns as well (holidays are lightly loaded and days after holidays are, as a rule, very heavy-loaded). For a reference on forecasting and modeling of ED arrivals, leading also to related literature, see Channouf et al. (2007). Internal Surgical and Orthopedic Figure 2 - Hourly arrival rates per patient type (averaged over 4 years) Arrivals in our simulation model are Poisson processes, with hourly rates that are forecasted for each future hour in question (say a shift, or a day) and each patient type. We use long term MA (Moving Averages) in order to predict hourly arrival rates. For example, in order to predict the arrival rate (assumed constant) on Tuesday during 11-12am, we average the corresponding arrival rates during the last 50 "Tuesdays 11-12am", excluding those that are holidays or days after holidays. The reason for choosing long-term MA is that we found it to provide essentially the same goodness-of-fit as more complicated time-series techniques. (Indeed, long-term MA, applied to the overall arrival rate over a test period of 60 weeks, gave rise to Mean Square Error (MSE) equal to 3.56, while two methods, based on Holt-Winters exponential smoothing, provide MSE=3.55 and 3.54). Another argument in favor of the use of long-term MA stems from the level of stochastic variability in historical samples, calculated for each hour-of-week, which fits 12

13 that of a Poisson process (Maman, Mandelbaum, Zeltyn 2009); then, the historical mean (or MA) is a natural (Maximum Likelihood) estimate for the Poisson parameter, namely the arrival rate. 4.4 Staff scheduling approaches With the present ED state assumed given (following Section 4.2), simulation is now to be used for predicting ED evolution, say several hours (a shift, a day) into the future; the goal is to determine appropriate staffing levels of resources nurses, physicians and support staff, as a function of time. Staffing the ED is a complex multi-objective problem. It must tradeoff conflicting objectives such as (1) Minimizing costs, (2) Maximizing resource utilization, (3) Minimizing waiting time of patients, (4) Maximizing quality of care. The complexity of such multi objective optimization, more so in a stochastic environment (e.g. randomness with respect to patients arrivals, routing, service durations, resources availability, and more) renders the optimization problem intractable analytically. This has thus led researchers to simulation-based heuristic solutions. A prerequisite for staffing is accurate forecasting of patients' arrivals, as described in Section 4.3. We then continue with predicting resource utilization; this leads to feasible staffing, based on pre-specified goals for resource utilizations (Section 4.4.1). However, the resources' view cannot accommodate the experience of patients for example, controlling the time until first encounter with a physician (Section 4.4.2). To control the latter, we calculate, for each resource type, its offered load as a function of time; then a classical staffing principle (square-root safety-staffing), in conjunction with the appropriate queueing model, yields our recommended time-varying staffing levels. In Section 4.5, a summary of our methodology will be presented Rough Cut Capacity Planning staff scheduling solution Rough Cut Capacity Planning (RCCP) is a technique for projecting resource requirements in a manufacturing or a service facility. As such, RCCP supports decisions regarding the acquisition and use of resources. Procedures for RCCP are listed in Vollmann, Berry, and Whybark (1993). These procedures are based on the estimated time on each product or service unit, and the allocation of the total time among the different resource types. The goal is to match offered capacity with the forecasted demand for the capacity of each resource type. Thus, RCCP algorithms translate forecasts into an aggregate capacity plan, taking into account the time each resource type spends on each type of product or service. We are proposing to apply RCCP in the ED environment, as follows: For each patient type i, calculate its average total time required from each resource r (e.g. physician, nurse): d ir. For each forecasted hourt, calculate the average number of external arrivals of patients of type i, Ai ( t ). Deduce the expected time required from each resource r at time t: RCCP ( ) ( ). r t Ai t dir The recommended number of units of resource r at time t, n r (RCCP,t), would be the load RCCP r (t), amplified by safety slack/staffing, or ST (we have used for in our experiments ST =90%): n r (RCCP,t) = RCCP r (t) / ST. We expect RCCP to achieve pre-planned resource utilization levels; its shortcoming, however, is that it ignores the time lag between arrival times of patients and actual times when these patients 13 i

14 receive service or treatment from ED resources. Since patients spend in ED several hours, on average, this time lag can be significant: the patients arrival rate frequently reaches maximum before the workload for a specific resource reaches maximum. This problem is remedied by our next approach Offered load approach The concept of offered-load is central for the analysis of operational performance. It is a refinement of RCCP in the sense that it spreads workload more accurately over time. For example, suppose that a nurse is required twice by a patient, once for injecting a medicine (10 minutes) and then, 3 hours later (in order to let the medicine take its effect), for testing the results (also 10 minutes). RCCP would "load" 20 minutes of nurse-work upon patient's arrival; the offered-load approach, in contrast, would acknowledge the 3 hours separation between the two 10-minutes requirements. Such time-sensitivity enables one to accommodate time-based performance measures, notably those reflecting the quality of care from the patients view point. In the simplest time-homogeneous steady-state case, when the system is characterized by a constant arrival rate and a constant service rate, the offered load is simply R = /= E(S) where E(S) is the average service time. The quantity R represents the amount of work, measured in time-units of service, which arrives to the system per (the same) time-unit. Staffing rules can be naturally expressed in the terms of the offered load: for example, the well known square-root staffing rule (Halfin and Whitt 1981; Borst, Mandelbaum, and Reiman 2004) postulates staffing according to n R R, (1) where >0 is a service-level parameter, which is set according to some Service Level Agreement (SLA) or goal. This rule gives rise to Quality and Efficiency-Driven (QED) operational performance, in the sense that it carefully balances high service quality with high utilization levels of resources. Arrival rates to an ED are, however, manifestly non-homogeneous and depend on the day-of-week and hour-of-day. Piecewise stationary approximations (such as SIPP - Stationary Independent Period by Period; Green, Kolesar, and Soares 2001), work fine if the arrival rate is slowly-varying with respect to the durations of services. This, however, does not happen in ED case. Assume that arrivals can be modeled by a non-homogeneous Poisson with arrival rate ( t), t 0. In this case, our definition of the offered load is based on the number of busy servers (equivalently served-customers), in a corresponding system with an infinite number of servers (Feldman et al. 2008). Specifically, any one of the following four representations gives it: t t R( t) E[ A( t) A( t S)] E[ ( t S )] [ ] ( ) e E S E ( ) ( ), u du ts u P S t u du (2) where A(t) is the cumulative number of arrivals up to time t, S is a (generic) service time, and S e is its so-called excess service time (See the review paper by Green, Kolesar, and Whitt (2007) for more details, as well as for useful approximations of (2)). Then, for calculating time-varying performance, we recommend to substitute (2) into the corresponding steady-state model, which is the classical M/M/n queue, or Erlang-C, in our case. To be concrete, assume that our service goal specifies a lower bound, to the fraction of patients that start service within T time units. Our QED approximation then gives rise to T t Rt t Rt 1 P{ W T} P{ W 0} P{ W T W 0} h( ) e, (3) q q q q t 14

15 Where h( t ) is the Halfin-Whitt function (Halfin and Whitt 1981). Equation (3) can now be solved numerically with respect to t, and the staffing rule (1) is replaced by the time-varying staffing function: n( OL, t) R( t) t R( t). (4) The above procedure has been called the "modified offered load approximations" readers are referred to Feldman et al. (2008) for additional details and further references. Square-root staffing are mathematically justified by asymptotic analysis, as workload (and hence the number of servers) increase indefinitely. (The practical motivation was large telephone call centers.) However, ample experience (as well as recent research; e.g. Janssen, Van Leeuwaarden, and Zwart (2008)) demonstrate amazing levels of high accuracy, already for single-digit staffing levels. This renders the above staffing rule relevant for EDs, as well as other healthcare systems, where the number of servers is indeed single-digit. (For small systems, one could always apply exact Erlang-C formulae. And indeed, we tested these exact calculations against the QED approximations in our experiments below, and the results were essentially unaltered.) Summarizing, we apply the proposed offered-load approach via the following steps: First, we are running the simulation model with infinitely many resources (e.g. physicians, or nurses, or both). Second, for each resource r (e.g. physician or nurse) and each hour t, we calculate the number of busy resources (equals the total work required), and use this value as our estimate for the offered load R(t) for resource r at time t. (The final value of R(t) is calculated by averaging over simulation runs.) Finally, for each hour t we deduce a recommended staffing level n r (OL,t) via formulae (4) and (3). 4.5 Methodology for forecasting short-term future ED state Our simulation-based methodology for short-term forecasting of the ED state is as follows: 1. Initialize with the simulation-based estimate of the current ED state. 2. Use the average arrival rate, calculated from the long run MA, to generate stochastic arrivals in the simulation. 3. Simulate and collect data every hour, for 8 future hours, using infinite resources (nurses, doctors). 4. From step 3, calculate staffing recommendations n r (RCCP,t) and n r (OL,t) using RCCP and Offered Load methods, described in Sections and 4.4.2, respectively. 5. Run the simulation from the current ED state with the recommended staffing. 6. Calculate performance measures. The above can be repeated with existing staffing (in Step (5)), which enables to compare it against RCCP and Offered-Load staffing. 4.6 Simulation experiments We now apply our methodologies through simulation experiments. First, we demonstrate the ability of our simulation-based tool to estimate current ED state, using a database from an Israeli hospital (Section 4.6.1). For that, we randomly choose a month (August 2007) in the database, and compare the known number of patients in the system with the simulation's outcome (following Section 4.2). In the second experiment (Section 4.6.2), we use the ED state at a specific time (September 2 nd, 2007, 16:00) to predict 1-7 hours ahead. (The chosen day is a Sunday, which, in 15

16 Israel, is a busy day of the week, being the first day following the weekend.) We then conclude, in Section 4.6.3, with a comparison of some ED performance measures, using two alternative staffing methods (following methodology, developed in Section 4.4) Current state We ran 100 one-month long replications of each scenario, in order to compare our simulation results with the data from hospital's database. For each date and hour, we calculated the average number of patients over the simulation replication (Avg series in Figure 3), and the corresponding standard deviation (SD), an Upper Bound (UP = Avg SD), and a Lower Bound (LB = Avg SD). In Figure 3, we depict 4 days, chosen to test our methodology against the (actual) number of patients from the database (Wip-Work in progress). We chose two periods that are two days long, the last day of the weekend (Saturday in Israel) and the first working day of the next week (Sunday). (For example, DOW_7_4 at time axis stands for 4am on Saturday and DOW_1_16 denotes 4pm on Sunday.) Number of Patients DOW_7_24 DOW_7_4 DOW_7_8 DOW_7_12 DOW_7_16 DOW_7_20 DOW_1_24 DOW_1_4 DOW_1_8 DOW_1_12 DOW_1_16 Day of Week (DOW) & Hour DOW_1_20 DOW_7_24 DOW_7_4 DOW_7_8 DOW_7_12 DOW_7_16 DOW_7_20 DOW_1_24 DOW_1_4 DOW_1_8 DOW_1_12 DOW_1_16 DOW_1_20 Figure 3 Comparing the Database with the simulated ED current-state (Weekdays and Weekends) These days are typically the calmest and busiest in the week, respectively. Note that the night and early morning shifts (hours 1-10 in Figure 3) are not overloaded (see, for example, the utilization profiles during 09-10, in Table 1), and performance measures are then less accurate. However, once the ED becomes congested, the simulation does yield an accurate prediction of the number of patients in the ED. At all times, though, the accuracy of prediction varies from reasonable to good. Remark. A probable explanation for somewhat worse fit of the simulation during lightly loaded hours is the following. When the load is low, the staff has more time for activities that are not incorporated into our simulation (e.g. department meetings). In contrast, during heavy loaded periods, there is virtually no time for such activities and reality becomes consistent with the simulation. LB UB Wip Avg 16

17 4.6.2 Forecasting staffing level Next, we looked at performance measures in the near future, to see if there is a way to improve ED operations via staffing. We looked at the offered load of all the relevant resources: Internal physician (Ip), Surgical physician (Sp), Orthopedic physician (Op) and Nurses (Nu). For our example, we use ED data until 16:00 and then apply simulation to forecast each succeeding hour, until the end of the day. In Table 1, we display the ED state until 16:00, then continued with the simulation-based forecast; the staffing levels used in the simulation is the one exercised in our partner ED we refer to it as "the existing staffing", and it appears in Table 2, under n(current). Columns Ip, Sp, Op, and Nu list utilization levels of the respective staff. (For nurses, this accounts for the time devoted to patients care, and excluding administrative duties; Physicians are exempted from the latter.). #Beds and #Chairs represent the number of occupied beds and chairs, respectively; %W is the fraction of patients that are exposed to unsatisfactory care, which here is taken to be "physician's first encounter occurs later than 30 minutes after arrival to the ED". In Table 2, we display the following characteristics: ED existing staffing - n(current), the offered load level (as explained in Section 4.4.2) in Offered Load column, recommended staffing level based on the offered load (aiming to achieve %W< 0.25 hour) n(ol), the RCCP level (as explained in Section 4.4.1) RCCP Load columns, RCCP staffing recommendations aiming at less than 90% staff utilization n (RCCP). Table 1 Simulation performance measures current and forecasted (existing staffing) Hour Ip Sp Op Nu #Beds #Chairs %W % 1% 23% 55% % % 25% 59% 68% % % 59% 67% 72% % % 45% 81% 58% % % 68% 94% 71% % % 62% 76% 63% % % 51% 46% 51% % % 43% 41% 53% % % 58% 46% 57% % % 46% 52% 50% % % 64% 70% 58% % % 64% 75% 56% % % 46% 60% 45% % % 38% 51% 46% % Table 2 Staffing levels (present and recommended) n (Current) Offered Load N (OL) RCCP Load n (RCCP) Hour Ip Sp Op Nu Ip Sp Op Nu Ip Sp Op Nu Ip Sp Op Nu Ip Sp Op Nu

18 4.6.3 Short-term forecasting performance and staffing levels In Table 3, we record simulated performance, under staffing levels calculated via the Offered Load and RCCP methods. As anticipated, the offered-load method achieved good service quality: indeed, the fraction of patients getting to see a physician within their first half hour at the ED is typically less than half of those under RCCP, the latter being also more influenced by the changes in the arrival rate. RCCP of course yields good performance at the resource utilization column, all being near the 90% target (for the resources with staffing levels larger than 1-2). It is interesting to compare Table 3 (planned staffing) with Table 2 (existing staffing): the latter has obvious hours of under- and over-staffing while the formers' performance is rather stable. (For example, n(current) implies understaffing during and overstaffing for period.) Preplanned staffing, either for resource utilization (RCCP) or, better yet, patients' service level (offered load), clearly has its merit. Table 3 Simulation performance measures (using OL and RCCP) Performance measures using Performance measures using OL recommendation RCCP recommendation Hour Ip Sp Op N Bed Chair %W Ip Sp Op N Bed Chair %W % 38% 40% 58% % 90% 54% 60% 59% % % 33% 35% 67% % 82% 47% 65% 81% % % 49% 53% 76% % 80% 45% 69% 92% % % 48% 57% 80% % 72% 43% 79% 97% % % 52% 65% 71% % 68% 46% 85% 99% % % 49% 59% 75% % 55% 45% 89% 99% % % 45% 50% 73% % 63% 39% 87% 99% % 4.7 Integration with decision-support system in ED In order to provide to decision makers (e.g. ED department manager) access to our solution, integration between our simulation and a Decision Support System (DSS) should be performed. Collecting real-time data from various sources can give a snapshot of the current situation and by using the methodology above, such a system can provide predicted information based on the current ED state. Then this information can be presented to the decision maker in various ways. For example, the ED manager can use this system for trying to avoid a future possible lack of resources (e.g. physicians, beds, nurses, etc.). Based on the above methodology, we were able to develop a DSS that presents, in a graphical interface, several important measurements of current and predicted factors of the ED (see Figure 4 and Figure 5). In fact, input to our system originates from numerous data sources. For example, ED current state is based on information from a multitude of hospital IT systems such as the Admit Discharge Transfer (ADT) system, the Picture Archiving and Communication System (PACS), the Lab Order Reservation system and the Electronic Medical Records system. Yet these systems provide only minimal operational information such as start and end of an activity. In particular, no information on queue lengths or waiting times is available therefore raising the need for our simulation-based capabilities of ED state completion and prediction. In the future, an increasing number of data sources will provide more and more information about the current state of the ED. A very significant upgrade of data collection capabilities can be achieved by the in- 18

19 corporation of an RFID system (see Section 5) that will provide information about location of patients, physicians, equipment, etc. Figure 4 - Dashboard snapshot showing rooms occupancy Figure 5 - Predicted arrivals and physicians load The hospital IT system collects its information and presents it to the user as a set of indicators and parameters. To interact with this hospital system, we have designed InEDvance (Greenshpan et al. 2009): a decision support system that can record, process, simulate, and present event data that hospital IT systems record and send, along with current and forecasted performance measures. The InEDvance system comprises algorithms that assist the ED manager in planning resources allocation for the next several hours, in order to handle forecasted resource scarcity. In particular, InEDvance has, at its core, a simulation-based module that is fed (in real-time) data from the hospital IT systems and then, through simulation (as described above), identifies and presents patients flow bottlenecks (e.g. excessive lines at the X-Ray) and consequently alert ED management. The information arriving from the various IT systems generates a dashboard of past, present and predicted activities within the ED. We sample-demonstrate the use of such a dashboard by combining it with our ED simulator, and graphically presenting (potentially in real-time) information on the dashboard, using a graphical user interface. Figure 4 shows a snapshot of the dashboard that presents, in various ways, past and current occupancy of the different ED rooms. Figure 5 demonstrates a dashboard tab that could alert, based on calculated forecasting indicators, against predicted congestion and resource shortage. 5 SIMULATION-BASED MODELING FOR SCENARIO VALIDATION: VALUE ASSESSMENT OF RFID TECHNOLOGIES IMPACT 5.1 Research goals and description of technologies In this section, we consider the validation of RFID technology implementation in ED. It is obvious that the actual introduction of RFID technology is costly and demands thorough re-design of ED processes and IT system. Therefore, there is a strong need to estimate benefits and costs of such implementation in advance and using relatively inexpensive evaluation procedure. The 19

20 simulation based modeling is a natural answer to this challenge. In this section, we consider implementation of two alternative existing RFID technologies: WiFi (802.11) and short range passive RFID. WiFi (see Emory and Leslie 2005, for example) is currently the most standardized and usable indoor wireless communication technology. Simple location tracking mechanisms can be built on top of an existing WiFi infrastructure. WiFi is designed to cover wide areas such as the overall hospital campus; hence, it can provide wide location tracking capabilities. The location tracking precision of WiFi, on the other hand, is poor. Naïve implementation uses the tag only for access point association and hence provides only room level resolution. Such installations may have also difficulties in distinguishing locations within two adjacent hospital floors. WiFi is based on active tag communication hence provides continuous location tracking. Passive RFID systems, on the other hand, offer very accurate location tracking, as tags can be identified only within short distances from the reader. The limited coverage issue can be resolved via additional readers, and by placing readers in designated frequently-accessed spots such as doors, pathways, mobile medical equipments (e.g. ECG machine) and patient beds. A significant advantage of passive RFID system is low tag cost. Passive RFID tags are disposable and require little to no maintenance. Thus, widespread deployment is more likely because tags can be given to patients, caregivers, families and visitors with little significant additional cost. Tags within a Passive RFID tags can be identified only during the reading transaction itself, hence they do not render continuous location tracking and monitoring. 5.2 Required process changes As the first step, it is necessary to define how various components of the ED processes will change given the new data provided by RFID. In addition, it is necessary to define which measures, or metrics, will improve due to the process change(s). It is important to specify the metrics that are expected to improve since only through these quantitative metrics, can the value of the RFID system be estimated (or the values of several RFID alternatives be compared) see Section 5.3. In general, there are three different types of metrics: clinical, operational and financial. In this research, we explore operational metrics that measure the operational efficiency of the ED and emphasize their relation to clinical and financial metrics. Average Length of Stay (ALOS) is an important example of an operational metrics. ALOS is the average amount of time a patient spends in the ED before either being released from the hospital or being admitted to a ward; one could account separately for patients who "left" due to other reasons, for example death or LWBS (see Fernandes, Price and Christenson 1997). Another important operational metric is workload - the average amount of work-time required from the staff, or a subset of it (nurses, physicians), quantified as a function of time. Note that the three above-mentioned types of metrics are interdependent. For example, if a patient waits for a long time before first examination by a physician, this may adversely affect an operational outcome such as ALOS which, in turn, could result in clinical deterioration, hence increased workload (more care required by the staff), and additional costs. For concreteness and demonstration purposes, we have chosen three ED processes whose importance for our hospital was established, for assessing the value of their improvements: From the operational point of view, implementing an alerting RFID system will help reduce unnecessary waiting times. We focus on patients who are "forgotten" in two Imaging areas: (a) in a remote CT area after completing their scan. Based on practice, we are assuming that 20

21 25% of such patients experience an average of one hour waiting before returning to the ED, when compared against an average of 10 minutes for regular waits. (b) the patients that are waiting after an X-Ray scan. Here "forgotten" patients wait just half an hour instead of the regular 10 minutes. (The X-Ray is relatively close to the ED and easier to locate "forgotten" patients at.) From the financial point of view, using patients RFID prevents abandonments of unregistered patients, and thus increases ED's turnover rate and, in turn, enhances hospital income. In essence, we measure the operational metrics LWBS fraction, which in its turn can help to determine the increase of hospital income. From the clinical point of view, using staff (nurses, physicians) RFID exposes physical layout problems, such as poor placement of rooms or equipment in the ED, which have adverse clinical consequences. Again, a related operational metrics staff walking distance, is considered. Excessive walking distances would indicate physical layout problems. 5.3 Simulation experiments To evaluate the benefits of using an RFID system for our three example processes, we have used an ED simulation model, described in Section 0, and programmed it to process six types of patients: Orthopedic, Surgical, and Internal, each in two conditions Walking and Acute (those in need of a bed). In addition, we made changes to the simulation in order to accommodate the expected impact of the two RFID technologies that are tested. For the process improvement, based on tracking abandonment, we made the following assumptions: Since data of actual abandonment times is currently unavailable, we distributed 4% abandonment (historical average for LWBS fraction) over five process steps: (1) waiting for a nurse to take patients anamnesis; (2) waiting for a physician's initial diagnosis; (3) after the physician's first examination and before sending additional tests; (4) while waiting for a physician to collect all the relevant data for further evaluation; (5) after further evaluation, while waiting to be released, hospitalized or for additional intensive tests. We assumed that WiFi technology identifies 100% of the abandonments and feeds those patients back into the process. Passive RFID, on the other hand, succeeds in only 50% of the cases. The difference arises because some patients would not abandon with their tags, while others might use vehicles, just as an example, to circumvent the passive sensors near the gates, which otherwise would detect them. Abandoning patients are not included in calculating lengths of stay, and they are naturally excluded from those who contribute to hospital profit. For the process improvement, dealing with reducing waiting times in the Imaging (CT or X-Ray) wards, we made the following assumptions and modifications: CT patients are waiting to return to the ED. Return timed is within 10 minutes for 75% of the patients and an hour for the rest. Passive technology is more effective than WiFi in this case: Passive technology accurately tracks room relocations and, therefore, gives rise to 100% reduction of the waiting time to 10 minutes. WiFi, on the other hand, reduces waiting times of only 50% of those who are expecting prolonged 60 minutes waiting. Of the delayed X-Ray patients, 20%, on average, are waiting 10 minutes and the others 30 minutes. 21

22 The Passive and WiFi systems were compared against two additional scenarios: an "ideal RFID system" and the prevailing situation without RFID. An ideal RFID system combines benefits of the Passive and WiFi systems: it succeeds to identify 100% of abandonments and tracks all customers who are forgotten at CT and X-Ray. Since the RFID influence on the two processes is intertwined (less abandonment can imply larger workload and waits), we decided also to check the influence of two changes (reduction of waiting times and preventing LWBS) separately. One week was used for simulation warm-up and three months of data for analysis, eleven simulation runs were performed for each case. Table 4 provides us with a summary of simulation results. It includes simulation averages for overall number of patients and LWBS patients, ALOS estimate, standard deviation of ALOS estimate (ALOS), based on variability between 11 simulation runs, and finally, (LOS) standard deviation of individual customer LOS. We shall analyze Table 4 data from several points of view. Average Length of Stay. Comparing the first three lines of Table 4, we observe that ALOS decreases once we reduce waiting times in the Imaging Units. Consistently with the story above, Passive RFID technology implies more significant improvement than WiFi. In contrast, LWBS reduction or elimination increases ALOS. Since patients are fed back into the process, congestion increases to a certain extent. (Garnett, Mandelbaum and Reiman, 2002, analyze such operational consequences of abandonments.) Finally, given the full implementation of RFID solution (Imaging waiting decreases and LWBS is reduced), Passive RFID provides ALOS that is slightly smaller with respect to the basic state, while WiFi implementation leads to ALOS increase. Table 4 The simulation results: comparison of different RFID systems RFID Number of patients (3 months) (3 months) LWBS System ALOS (ALOS) (LOS) Without RFID 24, (3.9%) WiFi, wait reduced 24, (4.0%) Passive RFID, wait reduced 24, (4.0%) WiFi, LWBS eliminated 23, (2.0%) Passive RFID, LWBS reduced 24, WiFi 23, (2.0%) Passive RFID 24, Ideal RFID (Passive + WiFi) 24, Number of LWBS patients. Table 4 shows that, in our simulation experiment, RFID technology fed back significant number of patients into the process. It should have both positive clinical and financial impacts: LWBS patients often return to an ED when their condition deteriorates; we also block attempts to leave the ED without providing payment guarantees. Another operational aspect of RFID implementation is captured by the intra-day staff workload, displayed in Figure 6. We calculate the workload in order to check that implementation of RFID will not lead to any unexpected operational phenomena. (Physicians that treat Internal Acute patients are chosen for this example). We observe that differences between RFID scenarios are not too large. 22

23 Workload Hour without RFID Passive, wait reduced WiFi, wait reduced Passive, LWBS reduced WiFi, LWBS eliminated Passive RFID WiFi RFID Ideal RFID Figure 6. Internal Acute physician workload Another dimension that we checked is the physical layout of the ED. From the simulation, we found that orthopedic physicians are walking about 2 kilometers per shift, between the walkingpatients area and the acute area (most times, there is just one orthopedic physician available for both locations. A second one would join from the orthopedic ward, when needed). Further investigation revealed that the distance between the two locations was excessive (about 100 meters) and the hospital managers took it into account in a redesigned ED. (See also Section 6.4 where the issue of excessive walking distances is discussed.) With the distance being that long, both WiFi and Passive systems identified (and could quantify) this problem easily. (WiFi, however, is less appropriate for measuring short-distance movements.) Considering all three aspects (clinical, economical, operational), which RFID solution should the manager implement? In our case, Passive RFID technology seems to be a reasonable option: it does not increase overall ALOS and prevents significant number of customers from abandonment. In addition, it is much less expensive than the WiFi implementation. However, in general, there is no clear-cut answer to this question. A decision-maker should take into account simulation results (especially ALOS, bed utilization and LWBS), implementation costs of different solutions, revenue from abandonment blocking, hospital preferences etc. In the ideal case, the cost/revenue optimization problem should be solved. However, it is not always easy to quantify financial impact of ALOS decrease or increase, often the impact depends on the staffing-level changes that can be implemented due to the change in the workload. Our simulation model can help to answer such sort of questions. 5.4 Integration with decision-support system: RFID-based control views The contribution of an RFID system to a hospital's environment should encompass two main aspects. First, RFID should have impact on daily routine and hospital staff; second, long-term im- 23

24 pact for strategic planning is desirable. Both aspects are implemented in the Decision Support System, introduced in Section 4.7. The system was designed on an IBM Cognos BI. Examples of interfaces with the processes in Section 5.2 will be now demonstrated. Online View in Figure 7 supports real-time decisions by hospital staff and executives depicting detailed events of hospital processes. These events contain information about specific patients, staff and services provided by the hospital. For our demonstration, we used our main discrete-event simulator. Figure 7a demonstrates how such an online view alerts on extreme waiting times of patients after CT services (the first process discussed in Section 5.2). Figure 7b demonstrates how a decision-maker is alerted on the presence of patients who attempt to abandon the ED (the second process discussed in Section 5.2), together with details of the process they have undergone until their abandonment attempt. Figure 7. Online view showing: a) patients waiting time for CT services b) patient abandonment The second Offline View in Figure 8 should be used for supporting long term planning. Therefore, it shows high-level details, aggregated over a pre-specified horizon. This view is to be used for high-level understanding and analysis of hospital processes, for example staff workload, quality of service, impact of decision-making and planning etc. Figure 8a displays patterns of patients arrivals rate over hours of a day and along days of week. It also highlights the magnitude of the gradient, thus pointing at the times of day when pattern-changes are the most significant. In such a view, we display averages over a year, which are to be used for planning and assessment of strategic and longer run tactical decisions. Figure 8b depicts workload on physicians at the hospital, through the analysis of patients waiting time for service excessive waits could trigger an alert. Figure 8. Offline view showing: a) Averaged patient arrival counts during daytime, in each of the weekdays b) Averaged patient wait time for physician 24