Social Science & Medicine 49 (1999) 1551±1566
www.elsevier.com/locate/socscimed
Assessment of ambulance response performance using a geographic information system Jeremy Peters a,*, G. Brent Hall b MapInfo Corporation, Troy, New York, USA Faculty of Environmental Studies, University of Waterloo, Waterloo, Ontario N2L 3G1, Canada a
b
Abstract The accessibility, distribution and utilisation of emergency medical services are important components of health care delivery. The impact of these services on well-being is heightened by the fact that ambulance resources must respond in a reliable and timely manner to emergency calls from demand areas. However, many factors, such as the unavailability of an ambulance at a center closest to a call, can adversely in¯uence response time. This paper discusses the design and implementation of a framework developed in a Geographic Information System for assessing ambulance response performance. A case study of ambulance response in three communities in Southern Ontario, Canada is presented that allows easy and rapid identi®cation of anomalous calls that may adversely aect overall operating performance evaluation. Extensions of the framework into a fully ¯edged service deployment and planning decision support system are discussed. 1999 Elsevier Science Ltd. All rights reserved.
#
Keywords: Ambulance response performance; Response anomalies; Geographic information system; Planning decision support
Introduction Emergency medical services (EMS) have a great impact on health service provision. Their importance is heightened by the fact that failure of ambulances to respond within set times to emergency calls from service areas may result in loss of life. When ambulances are not available at the center closest to a call or other unanticipated factors come into play, response time becomes a random variable, causing anomalies and considerable variation in response time performance. Hence, one of the primary objectives of the provision of emergency ambulance services is to deploy a limited number of vehicles in a way that ensures response time
* Corresponding author. E-mail address:
[email protected] (J. Peters) 0277-9536/99/$ - see front matter PII: S 0 2 7 7 - 9 5 3 6 ( 9 9 ) 0 0 2 4 8 - 8
standards are met, especially in life threatening and urgent call situations. Geographic Information Systems (GIS) technology is beginning to be used by health agencies in the planning of EMS deployment. GIS provides EMS planners with the ability to organise and manipulate large volumes of spatially referenced call data and to communicate spatial concepts to decision makers responsible for service deployment planning. Using GIS, decision makers are able to visualise data in map form and understand geographic patterns and trends in ambulance response performance that would otherwise be dicult to ascertain. Despite the advantages oered by the use of GIS in assessing ambulance service performance, the current body of research on this topic is limited both in terms of the number of applications reported in the literature and in their coverage of three important spatio-temporal dimensions of response time
# 1999 Elsevier Science Ltd. All rights reserved.
1552
J. Peters, G.B. Hall / Social Science & Medicine 49 (1999) 1551±1566
patterns and trends. These dimensions are the result of random variables, such as ambulance travel time and ambulance availability, that are often characterised by anomalies and considerable variation in response performance. The ®rst spatio-temporal dimension of emergency vehicle response time patterns and trends is the analysis and visualisation of response time anomalies and the `normal' variation in performance levels. Within this dimension, ambulance center location and vehicle deployment decisions should be based on well-de®ned areas of normal or consistent performance levels and these can only be clearly visualised if anomaly response times are extracted from all responses and treated separately. The second dimension comprises the appropriate use of a complementary set of response time performance indicators to evaluate trends in ambulance performance over space, time and by type of incident. The third dimension is the use of procedures to help explain aberrations in performance indicator patterns and trends. In order to inform better deployment decisions it is important to understand why an ambulance service has not met performance standards in a particular area. In many cases, seemingly unusual spatial patterns and trends in performance can be explained by analysing the variation in response times and examining the variables that aect this.
Objectives This paper presents a GIS-based application for improving the visualisation and analysis of ambulance service performance that employs an easy-to-use and robust methodology built around the three spatio-temporal dimensions of ambulance response patterns and trends noted above. The speci®c objectives of the paper are: 1. to present an analytic model for evaluating and improving EMS vehicle response; 2. from this, to design and implement a valid GISbased framework to assist planners in mapping and assessing EMS vehicle response; 3. to demonstrate the usefulness of the approach using empirical data, from the Ministry of Health in Ontario, Canada (OMH) in a case study.
Structure The Review section places the research within the modest current body of literature related to the use of GIS technology in EMS delivery. An analytic model and GIS design framework for evaluating and improving EMS vehicle response are then presented. The Data and operationalisation section describes the
research methodology, including the case study data and analytic processes programmed into the GIS. The Results section presents and discusses the results of an application of the approach for life threatening and urgent call data supplied by the OMH. The ®nal section concludes the paper with a brief discussion of the research contribution and a statement of future work required to produce a fully ¯edged GIS-based decision support system for EMS service deployment and planning.
Review GIS-based approaches have been adopted successfully to analyse the quality and timeliness of a range of emergency services including ambulance, ®re, and police dispatching, logistics, tracking and routing applications (Ward, 1994, p. 36±37; Barry, 1991, p. 74± 77; Gamble-Risley, 1997, p. 64±68; Bridgehouse, 1993, p. 611±622; GIS Newslink, 1993, p. 11; GIS Newslink, 1994, p. 13). The systems integrate a variety of associated technologies including Global Positioning Systems (GPS), Automatic Vehicle Location (AVL), ComputerAided Dispatch (CAD), routing algorithms, electronic maps and in-vehicle navigation to provide real time tracking, dispatching and routing of emergency vehicles. With these hybrid systems, dispatch managers can, for example, use AVL technology to track the real time locations of an ambulance ¯eet through GPS transponders attached to the vehicles and display their locations on GIS-based computer maps at the dispatch facility. Further, GIS is used in CAD to locate the addresses of calls on a geo-coded street network or property database and as a decision support tool to determine the optimal unit and route that should be taken to respond to each call. Automated EMS systems are used in jurisdictions world-wide, however there is relatively little published discussion of them. In one of the few reported case studies, a real time ¯eet management system located at the Sunstar Communication Center in Pinellas County, Florida is used to manage EMS response (Badillo, 1993, p. 161±176). This system geo-codes and displays on a wall-sized, color-coded, digital map the addresses of 911 emergency callers. The digital map also tracks and displays the location, heading, direction, and status of each vehicle in the EMS ¯eet. The system uses current information on the location, type, and status of each vehicle to identify the optimal unit to respond to each call. Once the optimal vehicle is identi®ed, the communication center dispatcher transmits a signal to the emergency vehicle's on-board computer to display a map of the surrounding area showing the vehicle's current location, the location of the emergency, and the best route to that location. The computer display
J. Peters, G.B. Hall / Social Science & Medicine 49 (1999) 1551±1566
1553
Fig. 1. An analytic model to assess ambulance response time performance.
also shows pertinent information below the map, such as the patient's name, nature of the injury, and whether lights and siren are required. The vehicle driver then transmits a signal back to notify the communication center that their ambulance has taken the call. While fully integrated EMS facilities such as the Sunstar Center are relatively rare, GIS is used more routinely by public EMS agencies. For example, the OMH Emergency Health Services Branch use GIS software to analyse the spatio-temporal patterns of historical call data to assess ambulance response and help plan service deployment based on these historical patterns. Also, the London Ambulance Service (LAS), in the United Kingdom, uses GIS-based desktop mapping software for the same purposes. Their application cal-
culates and maps the number of calls made in each ambulance center's territory. A statistical model is used to calculate how many calls are expected in an area (potential demand) compared with how many calls actually take place (realised demand). This application also analyses ambulance response by mapping the number of calls that exceed the nation-wide target of 14 min to reach the location of an emergency. A review of this application concludes that ``since the installation of the new control room systems, 95% of the calls were reached within 14 min, compared with 70% last year.'' (Super Solution, 1997, p. 16). Despite the advantages oered by the use of GIS in evaluating and improving EMS performance, no single application reported in the literature fully examines the three spatio-temporal dimensions noted in this section
1554
J. Peters, G.B. Hall / Social Science & Medicine 49 (1999) 1551±1566
to assess ambulance performance. An approach that incorporates these factors in a GIS framework to analyse ambulance performance is now discussed.
Analytic model The model shown in Fig. 1 uses a set of ®ve attributes for ambulance call data to assess response performance. These attributes include the number of calls within a time interval (such as a work shift or per week), the lapsed response time for each call, the purpose of each call, the date and time when each call is received, and the center location of the ambulance that responds to a call. Information availability can vary considerably between ambulance services in terms of the type and quantity of data that are collected and available for analysis. Hence, the intent of the model in Fig. 1 is to demonstrate how ambulance call data can be used more eectively, assuming that a data set comprising at least these ®ve attributes is available. Further, it is important to note that the central variable in this model is vehicle response time. This variable is simply the number of seconds, converted into minutes and seconds, that it takes for an ambulance, from the time of the receipt of the call at the dispatch center, to reach the call location. Although this variable measures time, it implicitly incorporates the eective time±distance for a response trip. Hence, it captures both the spatial and temporal dimensions of response that can account for the eect of numerous factors such as peak and o-peak trac ¯ows, weather conditions, and characteristics of the local street network on response performance. The ®rst dimension in Fig. 1 is the analysis and visualisation of response time anomalies and the `normal' variation in ambulance performance levels. This dimension is important, as areas where response times need to be improved can only be clearly visualised if anomalous calls are removed from the analysis to reveal the `normal' pattern of response performance. Moreover, it is important to separate anomalous response times from normal response times in order to help understand the cause, spatial distribution and statistical signi®cance of the anomalies and to predict the likelihood of their occurrence in the future. Anomalous responses can be identi®ed statistically by normalising the response time variable into standard z-scores and ®ltering cases that are outside 95% statistical probability, assuming a normal distribution, of the mean response time. Pre-processing the response time variable allows consistent performance indicators to be calculated in the next stage of the analytic model. The second dimension in Fig. 1 de®nes a complementary set of response indicators to evaluate trends in ambulance performance over space, time and by
type of incident. As noted above, average response time indicators should be calculated based on the `normal' variation in performance levels so that anomalies do not skew the averages and distort the spatial representation of ambulance response performance. Moreover, a spatial performance indicator, using the percentage of calls within a response time standard, can be used to provide a second measure to compare with the average response time indicator's representation of performance. This indicator provides an eective measure for visualising performance in relation to an ambulance service's response time standards. A relative performance indicator can also be used to measure all calls together since it is not aected by anomalous response times. In addition to using these two indicators to identify areas of acceptable and unacceptable performance for a given time period, it is also important to have a response time indicator that can evaluate trends in ambulance performance over time and by type of call. Planners should know which of these indicators have improved, remained constant, or worsened over time or by type of incident to help identify areas and time periods where ambulance performance needs to be improved. A performance indicator that shows the statistical change in average response times for two or more time periods or types of calls can provide an eective measure of trends in ambulance performance. Such an indicator allows several statistical hypotheses to be tested, including, for example, whether or not response times have signi®cantly improved after service modi®cation; whether response is better on weekends than week days; is worse during rush hour than during the rest of the day; or whether response is signi®cantly faster for life-threatening than for less serious types of calls. GIS-based spatial analysis can be used to subtract and compare two or more average response times for service areas. In addition, statistical analysis, using analysis of variance (ANOVA), can be used to determine the probability of a meaningful statistical dierence between two or more comparisons. The third dimension in Fig. 1 addresses the question of whether or not an ambulance service meets performance standards in a particular area. In many cases, seemingly unusual spatial patterns and trends in performance can be explained by analysing the variation in response time and the variables that aect this. For instance, frequency distribution tables can be generated to compare the frequency of unique response times and the variation in those response times represented by z-scores. The frequency of and variation in response times can help explain why a performance indicator value for one area is dierent from neighbouring areas. Further, variables such as the location of an ambu-
J. Peters, G.B. Hall / Social Science & Medicine 49 (1999) 1551±1566
1555
Fig. 2. A GIS framework to evaluate ambulance response performance.
lance responding to a call, the route taken, trac ¯ow, time of day, season/weather, and road conditions and restrictions can all cause anomalies and sometimes considerable variation in response times. These variables impact on response times by determining the distance and speed at which ambulances travel to respond to calls. The location of an ambulance at the time of response and its route taken are particularly important in helping to explain performance indicator patterns and trends because these variables are directly controlled and planned for by ambulance services. The other variables can be manipulated to varying degrees through resource deployment and route planning. As noted above, information availability can vary considerably between ambulance services in terms of their capacity to identify the location of an ambulance at the time of a call and the route taken in response. At a minimum, the center that the ambulance is based at can be identi®ed to help explain the impact of travel distance (center origin and call location) on response times. Using this information, territories `normally' serviced by each ambulance center can be de®ned and mapped. Frequency distribution tables can then be generated to compare the volume of calls answered by each center and the average response times for those calls in any given size area within each territory. Using these methods, it is possible to depict relatively easily using GIS, the Euclidean distances or time-distances travelled by ambulances based at each center in relation to their performance levels and the distances and performance levels of other centers and ambu-
lances. With the integration of appropriate statistical and geographical modeling methods and outputs into a GIS application that examines these spatio-temporal dimensions, historical call data can be used to assess more accurately ambulance response performance over time.
GIS framework A GIS-based framework for implementing centerbased spatio-temporal performance indicators is presented in Fig. 2. This framework describes the highlevel operational GIS design for the analytic model discussed in the Analytic model section. The framework describes the interaction between 11 components, including ambulance call data, spatial data, the system user, a graphical user interface, statistical analysis, spatial models, performance indicator calculations, built-in GIS functionality, mapping, evaluation, and decision making. Within this framework, an intuitive and easily navigable graphical user interface (GUI) allows EMS planners to interact with all features of the GIS environment. The GUI is a conceptual link between the user's interaction with the computer-based application and what it oers them as a decision support tool. It includes all of the considerations that EMS planners require to understand the application and to communicate eectively with it (Medyckyj-Scott and Hearnshaw, 1993). The GUI is also task-oriented, in that it enables users to generate performance indicator
1556
J. Peters, G.B. Hall / Social Science & Medicine 49 (1999) 1551±1566
maps by specifying the parameters they need to customise their output without having to deal with a complex series of native GIS functions. The GUI is the only part of the application that is visible and with which users interact. Thus, for the EMS planner, it is `the system' and its usability is of crucial importance. The application uses the parameters speci®ed in the GUI to evaluate ambulance performance through the use of appropriate statistical analysis and spatial modeling methods and response-based spatial performance indicator maps. EMS response mapping is a fundamental component of the GIS framework as it aords planners and decision makers the opportunity to preprocess data, identify and ®lter anomaly responses and to visualise quickly and easily problem areas which are performing below operating standards. Once problem cases are visually identi®ed, further evaluation via the statistical analysis and spatial model components allows understanding of the problem(s) to be re®ned and recti®ed. Statistical analysis and GIS database manipulation, spatial analysis and mapping functionality are programmed into the application to operationalise each component. For example, Structured Query Language (SQL) is used to manipulate and help transform ambulance call data. SQL is used for many tasks in the system, such as selecting subsets of call data based on parameters speci®ed by the user. SQL is used to transform call data into performance indicators, demand indicators, response time variation indicators (e.g. zscores) and frequency distribution tables. In order to accomplish these tasks, SQL uses aggregate functions to calculate averages and count records as well as data modi®cation requests (e.g. update) to transform data such as response times into z-scores using a customised algorithm. In addition to GIS database manipulation, spatial and statistical analyses are used to transform call data into performance indicators and to generate service territories or catchments. Spatial overlay analysis is a GIS function that compares map features and attributes across two or more map layers. This is used to generate a single change in average response times from two dierent maps, each showing average response times for a dierent time snapshot or for dierent types of calls. Further, ANOVA is used to show instances where the dierence in average response times is statistically signi®cant. Feature generalisation is an additional GIS function that determines contiguous groupings of identically valued map features and attributes which make an underlying pattern more readily apparent. This tool is used to generate functional aggregate service territories by combining individual service areas where a center has historically responded to the majority of calls. The interactions between the model, GUI, data, statistical analysis, and
GIS components of the framework operationalise the calculation and mapping of each performance indicator. In the evaluation component of the GIS framework, the spatial performance indicator outputs are used to evaluate a range of EMS response scenarios. EMS planners can assess the historical spatial distribution of service supply relative to demand. Further, planners can also determine the underlying patterns of accessibility in relation to emergency service response standards and identify areas with de®cient service provision (i.e. substandard response times). Moreover, they can determine the time and nature of the speci®c calls that have a signi®cant impact on response performance. Interactive tools are also used in the evaluation to generate frequency distribution tables to help explain performance indicator patterns and response time anomalies. This allows planners to determine, for example, if unsatisfactory responses are caused by responses from locations other than the closest center. Once planners determine which center has responded to a call, they can determine the pervasiveness of the problem in terms of the frequency and variation in response times in relation to response time standards. This information can help decision-makers plan service deployment in order to improve service provision by identifying areas where ambulance response is consistently outside the time standard. Using this information, decision makers can better target when, where, and for what type of calls ambulance response performance must be improved.
Data and operationalisation Data In order to test the analytic model and GIS framework described in the Review section, data from the Ministry of Health in Ontario (OMH), Canada were used. The OHM uses historical records of call data to represent the demand for and response to ambulance service calls. Until recently, the OMH produced two geographic reports, called `GeoPlot Reports', using manual cartography to show choropleth maps of the total number of calls and average response time for each cell in a 1 km2 Universal Transverse Mercator grid that covers the Province of Ontario. To achieve this, every emergency call is geo-referenced, as it is received, to its appropriate grid cell according to userde®ned periods of time (dates, day(s) of the week and time of day) and types of calls (e.g. Priority Code 4Ð life threatening). The GeoPlot reports have now been partially automated, using MapInfo GIS software. However, their usefulness still falls short of the needs
J. Peters, G.B. Hall / Social Science & Medicine 49 (1999) 1551±1566
1557
Fig. 3. Relationship between the ®ve processes used to operationalise the analytic model and GIS design framework.
expressed in the three dimensions of the analytic model in Fig. 1 and GIS framework described in Fig. 2. Five types of OMH historical call data were used to implement the procedures discussed in the Review section and to assess ambulance response performance. These data include the frequency of ambulance calls per center per 1 km2 UTM grid cell; the response time for each call (the elapsed time between the receipt of a request for an ambulance and its arrival on the scene); the purpose of each call (its OMH priority Code), the date and time when each call is received, and the center location of the ambulance that responded to the call (the center Code). The number of calls from each OMH 1 km2 grid cell provides an indicator of the level of EMS demand. This demand information is essential in helping to locate new ambulance services and to minimise response times in areas that have the greatest need for this service. As noted earlier, response time data can be transformed into spatial performance indicators to measure, for example, ambulance performance in relation to response time standards and the demand for service. The nature or purpose of a call indicates the type of demand and services required so that ambulance performance can be measured in relation to service standards and demand areas. For example, calls classi®ed as life-threatening generally have a shorter response time requirement and tend to occur in high potential demand areas, such as retirement communities, than those that are classi®ed as urgent but not life threatening. The date and time of a call is used to evaluate demand and performance in relation to a speci®ed time frame and also to establish daily trends in the receipt of calls. This information can be used to evalu-
ate, for example, whether there are discernible lulls in activity and whether there are repetitive peak times when life threatening health events occur (such as early in the morning or late at night). The location of an ambulance center provides a means of estimating the (straight line or city block) distance travelled to respond to a call. This information, in addition to being useful in planning future deployment strategies, helps explain ambulance response time anomalies.
Operationalisation Several methods were employed to operationalise the three dimensions and their relationships described in the analytic model and GIS design framework with the data outlined in the previous section. These methods are now described in terms of ®ve processes (Fig. 3). The ®rst process, response time variation and anomaly identi®cation, addresses the ®rst dimension in Fig. 1. This process enables calls to be treated as normal or anomalous for analysis within the second process, response time performance indicators. Three performance indicators are evaluated in the analytic model to address the second dimension in Fig. 1, namely response time threshold, average response time and dierence in average response time. The third (response time frequency distribution), fourth (call volume frequency distribution) and ®fth (service territory de®nition) processes in Fig. 3 address the third dimension in Fig. 1 and provide complementary methods to help explain performance indicator patterns and trends identi®ed within the second process. The service territory de®nition process requires call volume information generated in the call volume
J. Peters, G.B. Hall / Social Science & Medicine 49 (1999) 1551±1566
1558
frequency distribution process to generate territory polygons made up of adjacent 1 km2 grid cells where a single center has responded to a user-speci®ed minimum proportion of calls. The model used in the ®rst process in Fig. 3 de®nes two types of statistical anomalies, including single calls in a UTM grid cell and response times with a z-score of greater than 2.0 (i.e. response times to a grid cell that are signi®cantly dierent to `normal' times). This model provides the user with options to exclude calls with or without one or both types of anomaly from further analysis. A single call is considered an anomaly in this case because it is a poor indicator of a consistent performance level. The output from this process and ®nal step is an ambulance call database table. This table can be ®ltered to represent either all selected ambulance calls, selected calls without response time anomalies, or selected calls with response times that represent anomalies. If the user chooses the default option, the existing selected call data set becomes the default call table. If the user chooses the option to remove statistical response time anomalies from the selected call data set, then all call records that do not ®t the anomaly criteria are selected to create a call table without anomalies. Or, if the user chooses the option to map anomalies independent of all other response times, then all call records that meet the anomaly criteria are selected to create a table representing calls with anomalous response times. The resulting table enables anomaly and non-anomaly calls to be treated independently for the application of response time performance indicators. The dierence in average response time performance indicator (second process in Fig. 3) operationalises the second dimension of the analytic model to provide an eective measure of trends in ambulance performance over space, time and by type of incident. This indicator is calculated and thematically mapped to show the change in average response time for two dierent periods of time or types of calls. The actual model used for this performance indicator calculates the statistical dierence between two average response times using a one-way ANOVA. Negative change (e.g. ÿ2 min) in average response time minutes, shows an improvement in ambulance response, whereas positive change (e.g. +2 min) in average response time minutes, shows a decline in ambulance response time performance. The call volume frequency distribution process operationalises the third dimension of the analytic model to help explain performance indicator patterns and trends in terms of the distances ambulances based at each center travel in relation to their performance levels. This process generates a frequency distribution table showing the volume of calls responded to from
2
each ambulance center and the average response time from a center for any given area on a performance indicator thematic map. This process helps to explain the various geographical factors noted earlier and the service delivery system factors that intervene to either increase or decrease response time. The model used in this process de®nes an interactive performance indicator map, where the user can select a `call volume frequency distribution' button from the GUI and interactively query the graphic display of response times by selecting one or more grid cells. The frequency distribution of responses for all centers that responded to selected calls, are tabulated and displayed. This interactive model provides planners with a tool to obtain performance indicator patterns in terms of the ambulance centers that respond to emergency calls. The response time frequency distribution process also operationalises the third dimension of the analytic model to explain performance indicator patterns and trends. This process calculates the frequency of and variation in unique response times. The resulting table shows the frequency distribution of unique response times and displays the associated z-scores for each response time for any given area on a thematic map. Lastly, the service territory generation process operationalises the third dimension of the analytic model. This process allows the user to establish the functional catchments of ambulance service centers by examining the distances ambulances travel from their base center (in this case straight line, although other metrics could be added). Territories historically serviced by selected ambulance centers are de®ned and overlaid on thematic maps generated through the response time performance indicators process. The model used in this process de®nes an ambulance center's functional catchment as the area covered by adjacent grid cells where a center has dispatched an ambulance to respond to a user-speci®ed proportion of calls in each grid cell. For example, if 75% is speci®ed as the call threshold, then a center's catchment area is de®ned by all adjacent grid cells in which 75% or more of the calls were responded to by ambulances dispatched from that center. This approach allows a user to de®ne catchments dynamically, according to a user-speci®ed proportion of calls because there is no statistical or EMS standard for de®ning the percent of calls that comprise `normal' catchment areas.
Results Three purposively chosen areas in the Regional Municipality of Niagara in southern Ontario, Canada (Niagara Falls, St. Catharines and Welland) were used to test the analytic model (Fig. 1) and GIS framework
J. Peters, G.B. Hall / Social Science & Medicine 49 (1999) 1551±1566
1559
Fig. 4. Study areas and UTM grid cells (St. Catharines, Niagra Falls and Welland).
(Fig. 2). The areas are in close geographic proximity to each other, comprise a mixture of predominantly urban (Niagara Falls and St. Catharines) and predominantly rural (Welland) communities, each has at least one ambulance service, the services respond to each other's calls when an ambulance from the closest center is not available, and the OMH still assesses realised EMS response in these areas using manual mapping methods. The 1 km2 spatial reference UTM grid used by the OMH is employed to de®ne the spatial extents used for the three study areas (Fig. 4). Each ambulance center is uniquely identi®ed by its service area and center Code (e.g. service area 143 center 00 is Niagara Falls center (143 00)). Life threatening (Code 4, requiring 8 min or less response time) and urgent (Code 3) priority call response times for the period 1995 to 1996 inclusive comprise the test data. The response time variation and anomaly identi®cation process was applied to the ®rst dimension of Fig. 1. This component of the analytic framework provides the user of the GIS framework with options to exclude calls with one or both types of anomaly (e.g.
single calls in a grid cell and response times that have a z-score beyond 2.0) from further analysis. Fig. 5 shows the average response time accessibility surfaces for the three study areas before the response time variation and anomaly identi®cation process was applied and anomalous response times removed from the analysis. The average response time for all calls per cell (min and s, rounded to the nearest 10 s) is written to each cell and cells are thematically shaded into preset average response time classes for ease of interpretation. Inspection of Fig. 5 shows that, as expected, average response times are greatest at the periphery of the study areas, relative to the actual locations of the ambulance centers and there are numerous cells in each study area with average response times beyond the acceptable 8 min maximum for priority 4 calls. Without further analysis, the extent to which these patterns and response times are aected by anomalies in the call data remains unclear. After the removal of anomalous calls according to the criteria speci®ed above, Fig. 6 shows that the aver-
2
J. Peters, G.B. Hall / Social Science & Medicine 49 (1999) 1551±1566
1560
Fig. 5. Average response time performance indicator (all calls).
age response times in all three study areas are improved to be generally within OMH operating standards. In St. Catharines and Niagara Falls especially, large areas of 9 km2 and 5 km2 respectively change their status in terms of meeting the Ministry response standard of 8 min or less. Grid cells with more than 8 min average response time decrease in frequency in total from 48 to 32, from 28 grid cells to 18 in St. Catharines and from 12 to 7 cells in Niagara Falls. Also, two grid cells changed their status in Welland (one to only marginal acceptability). Hence, a clearer, more consistent, more acceptable and more realistic pattern of ambulance response is shown in all study areas after the removal of anomalous call data. Fig. 7 shows the number of calls per cell with response times beyond 2 z-scores of the mean response time for all calls. The frequency and pattern of these anomaly calls is similar to the pattern of total call frequencies, suggesting that response time anomalies impact on response performance roughly proportionately to overall EMS demand. On the peripheries of the study areas, where demand is lowest, there are
2
fewer total calls, fewer response time anomalies (Fig. 7) and overall response time variation is greatest (Fig. 5). In the immediate vicinity of the ambulance centers, overall demand is higher, there are more call outs and more response time anomalies, as de®ned (Fig. 7). This pattern is explained by the fact that calls to grid cells in the vicinity of the ambulance centers are often responded to by ambulances returning from other callouts, thus extending their call-speci®c response time beyond the normal variation, or by response swapping with an ambulance from a center other than the nearest to the location of the call-out. The results show that by applying the response time variation and anomaly identi®cation process, planners can better identify areas where ambulance response is consistently outside the expected response time standard. Once problem areas are identi®ed, their causes can be investigated and remedial plans implemented. This more reasonable representation of the underlying patterns of accessibility can therefore help EMS planners and decision-makers deploy services in order to improve service in the aected areas.
J. Peters, G.B. Hall / Social Science & Medicine 49 (1999) 1551±1566
1561
Fig. 6. Average response time performance indicator (anomalies removed).
The dierence in average response time performance indicator was used to address the second dimension of the analytic model and thereby provide a further evaluation of response time performance. This process provides EMS planners with an eective measure of dierence in ambulance performance over space, time and by type of call. Unlike the average and relative response time indicators, the results of this process show when and where the socio-spatial dierentiation (e.g. considering locational factors, demographic characteristics, and organisational characteristics of the EMS delivery system) of actual demand is a signi®cant factor in response performance in relation to the corresponding spatio-temporal patterns of realised response times. In Fig. 8, the socio-spatial dierentiation of response times shows a clear pattern. Grid cells where Code 4 calls have statistically signi®cantly lower average response times than Code 3 calls are distributed ubiquitously with the exception of a relatively small number of cells at the periphery of each study area, where there are a small number of calls and no signi®cant
dierence in average response time. Given the prioritisation in response urgency in Code 4 over Code 3 calls, very large dierences in average response times are generally evident in grid cells further from the ambulance centers. In fact, in one peripheral cell in Niagara Falls, Code 4 calls are responded to on average 14.5 min faster than Code 3 calls, although the range is generally between 2 and 5 min. The OMH analysis of the demand for and response of ambulance services focuses on these two types of calls, where speed of ambulance response and the deployment of services are most important. This pattern illustrates how organisational factors of ambulance service delivery, such as the prioritisation of requests for ambulance services, can have a potentially signi®cant impact on response time performance. Although not shown in Fig. 8, the dierence in response times based on day and night work-shifts is statistically signi®cant in more grid cells than not, in all three study areas with day shift calls having a lower average response time than night shift calls. This dierence is more than likely accountable for by night shift
1562
J. Peters, G.B. Hall / Social Science & Medicine 49 (1999) 1551±1566
Fig. 7. Anomaly call distributor (>
drivers not being as alert as day shift drivers and that night driving requires greater caution than driving during the day. Moreover, there may be fewer night shift drivers on duty, so the likelihood of multi-call trips increased and call response swapping may play a greater role than during the day. For the third dimension in Fig. 1, the response time frequency distribution, call volume frequency distribution and service territory generation processes are evaluated in terms of how the models help to explain response time anomalies and performance indicator patterns in the three study areas. This analysis allows EMS planners to determine, for example, the extent of
2 2 z-scores).
response call swapping and its impact on response time. Once planners determine which center is responsible for a call response, the response time frequency distribution process can be used to examine the frequency and variation in response times in relation to operating standards. The Niagara Falls ambulance service is used to illustrate this dimension of Fig. 1. In this case, one unusually high average response time of 10 min and 50 s (over ®ve calls) in the south west corner of the Niagara Falls response area (Fig. 5) was skewed by a single particularly poor response time belonging to an ambulance based in St. Catharines (center 220 00)
Table 1 Call frequency distribution for one cell in Niagara Falls Center identi®er
Number of calls
Percentage of calls
Average response
143 00 143 02 220 00
3 1 1
60 20 20
8:15 9:49 19:28
J. Peters, G.B. Hall / Social Science & Medicine 49 (1999) 1551±1566
1563
Fig. 8. Dierence in average response time (Codes 4 and 3).
responding to an out-of-area call (Tables 1 and 2). The functional catchment area for Niagara Falls (Fig. 9) shows that this is the only grid cell in the southern part of the study area where the Niagara Falls ambulance center (143 00) did not respond to 75% or more of the calls. In fact, Table 1 shows that the Niagara Falls center responded to 60% of the calls in this cell, with an average response time of 8 min and 15 s. This is 2 min and 35 s less than the average for all calls to this cell, and is consistent, albeit slightly higher, with
the pattern shown for the surrounding area (Figs. 5 and 6). The catchments and frequency distributions identify performance levels according to the various centers that responded to calls in areas where performance standards are not met. This allows planners to determine if unsatisfactory accessibility is caused as a result of calls that are responded to by centers other than the closest center. Use of the GIS framework in this context allows quick identi®cation of the cause of high
Table 2 Response time frequency distribution for one cell in Niagara Falls Response time (rounded min)
Number of calls
Percentage of calls
6 8 10 11 19
1 1 1 1 1
20 20 20 20 20
z-score
ÿ1 ÿ0.6 ÿ0.2
0 1.6
1564
J. Peters, G.B. Hall / Social Science & Medicine 49 (1999) 1551±1566
Fig. 9. Functional ambulance center catchments (>75% of all calls).
average response times through the integration of visual inspection of response map output and simple frequency analysis; all operationalised through the custom programmed GUI of the GIS. Table 3 contains a summary of the call volume frequency distribution process applied to all calls in the
three study areas. The table shows that calls responded to by ambulances from other areas (call-out response swapping) represent a much greater proportion of calls with anomalous response times (Fig. 7) than those with normal variation in response performance (Fig. 6). For example, in St. Catharines the percentage of
Table 3 Summary of the call volume frequency distribution tables for all three study areas
Number of calls by local centers (Fig. 6) Number of calls by other centers (Fig. 6) Number of calls by local centers (Fig. 7) Number of calls by other centers (Fig. 7) Percentage of calls by local centers (Fig. 6) Percentage of calls by other centers (Fig. 6) Percentage of calls by local centers (Fig. 7) Percentage of calls by other centers (Fig. 7) Average response time (Fig. 6) Average response time (Fig. 7
St. Catharines
Niagara Falls
Welland
Total
2878 520 70 52 84.7 15.3 57.4 42.6 6:34 15:12
1908 166 54 33 92 8 62.1 37.9 5:44 12:49
922 56 26 14 94.3 5.7 65 35 5:51 12:15
5708 742 150 99 88.5 11.5 60.2 39.8 6:11 13:54
J. Peters, G.B. Hall / Social Science & Medicine 49 (1999) 1551±1566 calls with anomalous response times that represent responses by ambulances from other areas (42.6%) is considerably higher than the percentage of calls with normal variations in performance that were responded to by ambulances from other areas (15.3%). Generally, the call volume frequency distribution process shows that when ambulances respond to calls from locations that are nearer to other centers, response time becomes a random variable causing anomalies and response performances outside those required by provincial operating standards. Before remedial actions can be taken through revised service planning to bring response times inside the required standards, it is necessary to understand the magnitude, the spatial extent and the eect on call response time of the response swapping problem. The processes available within the GIS to address the third dimension of Fig. 1 render this information to decision makers in both visual, map-based and statistical form.
Conclusion This paper has presented and operationalised an analytic model and GIS framework for assessing ambulance response performance. The model and framework were applied to ambulance response call data from three study areas in southern Ontario, Canada. Results were presented which show that the approach is successful in facilitating the visualisation and analysis of response time anomalies relative to `normal' variations in ambulance performance levels. The GIS framework is particularly useful in this context as it allows, through use of a custom programmed graphical user interface, an EMS deployment planner to produce quickly in map and tabular form easily interpretable outputs from quite complex source data. The patterns of response performance for any userspeci®ed priority type(s) and time(s) of ambulance callouts allow easy identi®cation of problem areas (i.e. where actual performance is unacceptable according to mandated response performance) for subsequent analysis. The identi®cation and exclusion of anomalous response times from further analysis is facilitated through the use of one or both of two ®lters. However, the GIS framework and analytic model are suciently open-ended to allow for the inclusion of additional ®lters. Exclusion of 1 km2 cells with a single call-out (since it is impossible to calculate any descriptive statistics from a single observation) and cells with an average response time beyond two standard z-scores, bought, for the most part, ambulance response performance for the study areas within OMH operating standards. The cells with anomalies as de®ned, although excluded from response performance analysis, are retained
1565
within the GIS framework for further scrutiny. It was shown that in many instances poor response performance can largely be explained by operational/organisational factors, such as too few ambulances at a center to respond to calls as they come in and response swapping that may necessitate responses from centers other than that which is closest to the location of a call. Hence, the GIS framework allows not only the distribution of call anomalies to be identi®ed and visualised but also it allows their causes to be established with reasonable certainty. To this extent, the use of the automated GIS approach for ambulance response analysis oers considerable advantages over manual mapping and pure database-oriented methods. Moreover, the integration of statistical analyses with automated map output from voluminous and often complex source data is particularly advantageous. Despite these advantages, there are a number of areas where the GIS approach presented in this paper could be expanded and improved to oer true decision support capability to EMS deployment planners and decision makers. A complete GIS-based EMS decision support system would require, in addition to the components described in this paper, the inclusion of methods to analyse factors (such as population density, population age, income, land values and land use mix) that in¯uence the socio-spatial dierentiation of ambulance service demand and models to evaluate the impact of service center locations on response performance. Further, measures of potential geographic accessibility could be incorporated into the analytic model and the GIS framework to model the potential for individual use behaviour in relation to a surrogate measure of need that accounts for the variation in actual response times. The development of a scenario-based probabilistic location±allocation component would also contribute to the emergency vehicle response planning process by determining the locational con®guration of facilities that best minimise system-wide or call/time speci®c response time and optimise the allocation of resources to centers. This knowledge would allow EMS planners and decision makers to know how to best allocate resources, such as the number of ambulances and shift drivers, to minimise optimal response time. Moreover, in combination with solving the resource allocation problem, the optimal center location solution would allow the role of current center locations on call response time to be assessed. Further, such `what if' scenario-based capabilities would allow decision makers to identify optimal new center locations and resource levels, given the current locations and resource levels at existing centers. This type of approach, grounded in GIS and utilising spatial data modeling can have substantial bene®ts for managing
1566
J. Peters, G.B. Hall / Social Science & Medicine 49 (1999) 1551±1566
and analysing data to produce information relevant to decision making and in simulating the eects of dierent planning decisions on EMS performance.
Acknowledgements The authors would like to express their thanks to two referees, whose suggestions signi®cantly enhanced the presentation of this paper.
References Badillo, A.S., 1993. Transportation and navigation. In: Castle, G.H. (Ed.), Pro®ting from a Geographic Information System, 3rd ed. GIS World Inc, Fort Collins, pp. 161±176.
Barry, D., 1991. Fleet management makes advances with digital mapping technology. GIS World 4 (7), 74±77. Bridgehouse, B., 1993. Emergency health services: a GIS application. In: Proceedings of the 5th Canadian Conference on GIS, Ottawa, March 23±25, pp. 611±622. Gamble-Risley, M., 1997. Emergency tools provide disaster relief. Government Technology 10 (8), 64±68. GIS Newslink, 1993. Emergency mapping response service implemented (Dauphin County, PA). GIS World 6 (7), 11. GIS Newslink, 1994. EIS implements real-time emergency tracking. GIS World 7 (12), 13. Medyckyj-Scott, D., Hearnshaw, H.M., 1993. Human Factors in Geographical Information Systems. Belhaven Press, London. Super Solution, 1997. Mapping ambulance priority. MapWorld Magazine 2 (1), 16. Ward, A., 1994. Saving lives in the West CountryÐusing GIS to improve ambulance response times. Mapping Awareness, April, 36±37.