Effectively predicting the demand for ambulances across regions can both improve the operational capacity of emergency services as well as reduced costs by optimizing resource utilization and providing an optimal spatial deployment and duty planning of paramedic crews. This results in quicker response times in attending emergency incidents reducing fatalities. Moreover, as ambulances play a critical role as first responders, calls to the ambulance service provide precious real time epidemiological information traces that can assist population health monitoring at scale and lead to improved policy design in healthcare.
Past studies aiming to explain geographic variations in the volume of calls for emergencies (jones2005circadian; carter2001scheduling; o2013system; peacock2006emergency) have been limited to examining epidemiological patterns across very broad geographic scales (national level). Enabling predictions at finer spatial scales, e.g. at the level of city neighborhoods, can generate intelligence that will allow the targeting of healthcare interventions in a more accurate manner, specializing treatment to the characteristics of populations in need. The importance of geography for health in fact has been highlighted through works pointing out that postal code may be a better predictor, compared to genetic information, when it comes to explaining the well being of local populations (graham2016your; chetty2016association).
In this work, our aim is to estimate the volume of ambulance calls at the level of individual Lower Super Output Areas (LSOA) in the North West of England. We investigate various environmental, socio-economic and demographic factors that can contribute to the rise of emergency incidents at a givenlocale. In this setting, we note how a population’s level of deprivation has a deterring impact on its health status with deprived urban areas being those where the number requests for emergency medical attention surges. Additionally, we identify regional population volume dynamics as one of the primary drivers for emergency calls to take place, showing how health incidents are likely to occur in areas where people become active and not simply those where they are registered as residents according to census. Critically to the novelty of the present work, we exploit place semantics and mobility patterns in location based-service Foursquare to infer population activity trends at local areas and attain more accurate predictions of the volume of calls an area will experience. Our research findings are described in more detail next:
Ambulance calls concentrate in urban areas and form patterns of spatial co-occurrence: in Section 3
we show how the spatial distribution of calls to emergency services is highly skewed, with a large fraction of activity being concentrated in major urban centers. Furthermore, there are strong patterns in terms of how incident types spatially co-occur. For instance,overdose/poisoning cases are highly correlated across geographies with convulsions/seizures and unconscious/fainting incidents. On the other hand breathing problems tend to correlated more with chest pain complaints and sick person cases.
Higher regional levels of deprivation imply higher volumes of ambulance calls: In Section 4.3 characterize geographic areas using various socio-economic indicators accessed through open government datasets, including we scores of geographic regions (noble_measuring_2006). We find that breathing problems, chest pain and psychiatric/suicide related incidents are more common in areas with higher crime rates and lower income levels.
Daytime population levels are a better predictor of ambulance calls than residential population: In Section 4.2 we define a variable to estimate daytime population levels. This is the sum of the number of workers at an area, younger (below 16 year old) and older (above 65 year old). We describe the stark differences between the spatial-distribution of daytime and residential populations showing how the former yields a much higher correlation score (pearson’s vs ) with the total number of calls in an area and is key to explaining variations for a set of incidents types including traumatic/injuries and uconscious/fainting.
Foursquare activity patterns at urban regions contribute to better predictions in ambulance calls for an area: Finally, in Section 5 we formulate a prediction task where our goal becomes to combine the various information sources discussed above and assess their relative importance in predicting the number of ambulance calls at an area. Daytime population levels are the most significant factor in explaining variations in ambulance calls, followed by the index of multiple deprivation for an area and foursquare check-in activity. The importance of each variable however, depends on the type of incident considered. Daytime population level best explain high number in falls and traumatic injuries whereas calls at areas with increased levels of unconscious/fainting and overdose/poisoning incidents are best approximated using check-in frequencies from location-based service Foursquare. The service proves to be a useful proxy of population activity at Food and Nightlife places.
Our work demonstrates how traditional, yet critical, sectors in healthcare may be improved through the integration of digital datasets from online sources. Location-based technologies could improve the operational efficiency of emergency services through more refined descriptions of population activities at fine geographic scales. From an epidemiological perspective they can be integrated with socio-economic and demographic indicators to offer a deeper understanding on population health patterns.
2. Related Work
|Lower Super Output Area Boundaries (data.gov.uk)||LSOA shapefile polygons|
|2011 Census; Table PHP01 (data.gov.uk)||#People residing in LSOA|
|Communal Population (CmmnlRs)||#People in communal establishments|
|Area size||LSOA area in hectares|
|Average Household Size (AvHshlS)||#persons in household (mean)|
|Workplace Population 2011111www.nomisweb.co.uk/census/2011/wp101ew||Employed People 16-74|
|LSOA Mid-Year Population Estimates 2011 (ons.gov.uk)||# Persons at each age in years|
|English Indices of Multiple Deprivation (www.gov.uk)||Deprivation scores for a local area|
|Foursquare check-in data||logged check-ins and place categories by LSOA|
Health Geography and environmental epidemiology
Spatial analysis and health geography trace their roots back to London when John Snow famously drew maps with markers of health incidents to locate the source of a cholera outbreak (snow1855mode). Spatial epidemiology has ever since contributed to our understanding of how diseases spread and appear geographically (gatrell1996spatial) tracing its roots on spatial statistics and quantitative geography (fotheringham2000quantitative). The state-of-the-art in statistical epidemiology of non-infectious disease usually treats occurrences as a spatio-temporal point process(moller1998log; diggle2007model; diggle2005point). This family of techniques however focuses on predicting incident frequencies of a single disease and does not provide interpretations on the environmental and demographic factors that may drive the spatio-temporal occurrence of medical incidents. They trace their roots in methods such as kriging (stein2012interpolation)
and they effectively reduce the problem of modeling the spatio-temporal occurrences of epidemiological incidents to a form of interpolation.
More recently, research literature in the field of health geography has focused on explaining geographic variations in emergency requests in terms of census and demographics data which has become available on a national level (ong2009geographic). Deprivation levels of urban communities (e.g. accessibility to employment) or differences between rural and urban areas have been projected to partially explain geographic variations in the volume of calls for emergencies (o2013system; peacock2006emergency). Our work considers urban deprivation factors in explaining call variations at a fine level of spatial granularity of areas with a few hundred residents. Moreover, we investigate the interplay of environmental, demographic and urban activity factors across different incident types (e.g. psychiatric, assault, fall etc.).
Social media in health analytics.
The field of digital health has risen in recent years thanks to the proliferation of the web as well as mobile sensing technologies (servia2017mobile). Despite their biases (lazer2014parable), web and online social media sources, provide ample opportunity to break away many of the barriers that characterize traditional experimental methodology in medical studies. These include being able to track users health behavior in a social context and anonymously (de2016discovering; de2014mental), or at large population scales (paul2016social) while retaining the benefits of fine spatio-temporal views on user behavior (mejova2015foodporn). An aspect of novelty in the present work regards the incorporation of information from social media services to understand ambulance demand regionally. Geo-referenced datasets from services like Foursquare have the advantage of providing us with place semantics and mobility patterns described at fine spatial scales.
3. Operational scope & data
3.1. The North West Ambulance Service
The North West Ambulance Service NHS Trust (NWAS)222http://www.nwas.nhs.uk/ is the second largest ambulance trust in England, providing services to a population of seven million people across a geographical area of approximately 5,400 square miles. The organisation provides a 24 hour, 365 days a year accident and emergency services to those in need of emergency medical treatment and transport, responding to hundreds of thousands emergency calls per year. Highly skilled staff provide life-saving care to patients in the community and take people to hospital or a place of care if needed. Calls that result in an ambulance dispatch may come via the (The Europe-wide also results in a call). The call operator will ask the caller a series of questions to ascertain the degree of emergency and will assign a dispatch code to the call. Dispatch codes are numeric and correspond to a broad classification of incidents (e.g. falls, traumatic injuries, assualt, psychiatric etc.). If the call requires a response, an appropriate team receives the instruction, and then swiftly makes its way to the incident location, using an onboard satelite navigation system.
We next describe the characteristics of the data employed in the present work. The primary source of data, the ambulance calls, has been collected by the North West Ambulance Service (NWAS). We employ numerous datasets to design a number of variables from demographic, socio-economic and web sources (Foursquare). These data layers will let us assess the efficacy of various information sources in predicting geographic variations in ambulance calls.
Ambulance calls dataset
The data provided by NWAS are those routinely collected as an emergency call operator receives a call where it is anticipated an ambulance may be required. The data is comprised of million calls the service has responded to from April 2013 to March 2017. Each incident has a dispatch code which corresponds to the type of the medical condition or cause that led to the call (e.g. suicide, fall, traffic incident etc). Codes range from 1 to 35 though other codes for rare cases may also be used. We exploit the incident number to stratify ambulance calls by nature and ask the question whether different incident types are associated with different factors. Critically to the present work, we exploit geographic information on where the incident took place at the administrative level of lower super output areas (LSOAs) which in the UK corresponds to the first four letters of the zip code. We explain LSOAs in detail next.
Spatial boundaries, populations, demographics and socio-economic indicators
Table 1 summarizes the additional data sets in terms of the variables that we utilise. The Lower Super Output Areas (LSOAs) are the fundamental unit of spatial aggregation considered in this work. Output areas were originally created such that populations were approximately similar socially and in size (cockings2011maintaining). LSOAs were assembled to maintain such similarity with a target population size of around 1500, but naturally, there is some variation as we demonstrate in the next section.
In terms of population data, we use information on the number of people residing in each LSOA, the number of people in communal establishments (communal population) and average household size. The workplace population corresponds to an enumeration of the people that work in an LSOA. In Section 4 we combine workplace population with residential population of young and older age groups (non working populations) to define the daytime population variable, which becomes one of the best predictors for ambulance calls.
In terms of socio-economic indicators, we employ The Index of Multiple Deprivation (IMD) which is a score calculated by the government in the UK to characterize areas through the consideration of a set of deprivation and quality of life indicators. These include income and crime levels, accessibility to education, health deprivation and disability, barriers to housing as well as the quality of the living environment. IMD has been shown to be a very important discriminative signal when aiming to predict the dynamics of complex urban processes including gentrification (hristova2016measuring). The number of healthcare providers corresponds to an enumaration of health services that are present in an area, which includes hospitals, general practioners (GPs) and social support facilities amongst others.
Finally, we employ a dataset of public Foursquare check-ins pushed on Twitter, collected over 11 months in 2011. For every check-in information about the place the user has checked-in becomes available, including its location in terms of latitude and longitude coordinates. Additionally, for every Foursquare venue we know the category of it (Coffee Shop, Italian Restaurant etc.). For the purposes of the present work we use the higher level Foursquare categories (Food, Nightlife, Travel & Transport, Residences, Arts & Entertainmens, Shops, College & University Outdoors & Recreation, Professional & Other places.)333https://developer.foursquare.com/docs/resources/categories. Overall in the area covered by the North West Ambulance service, we have observed approximately thousand check-ins over the month period considered. We observe significant practical correlations between the Foursquare and ambulance call datasets despite the the fact that their time windows do not overlap (2011 versus 2013-2017). Finally, we have associated every Foursquare venue with an LSOA through a spatial join of the venues dataset with the polygons describing the boundaries of the LSOAs.
3.3. Spatial distribution of calls
The main focus of the present work is to predict the geographic variations of ambulance calls and the identification of the factors that drive their increase. To highlight the relevance of the question, we provide views on how ambulance calls are skewed across different geographic regions (LSOAs) in the North West of England. The choropleth map for the total number of calls across the whole four year period is shown in Figure 1, using the natural logarithm to represent the number of calls in each LSOA. One can see that the sizes of the LSOAs vary considerably to maintain approximately similar population sizes, with urban centres corresponding to smaller geographic areas of higher population density. The squares on the main map show areas zoomed in to reveal more detail in the two main urban centers of the region, Manchester and Liverpool. As can be observed, there is a significant variation in calls across areas. Table 2 shows the ten of those LSOAs with the highest numbers of calls. The highest rated is the LSOA with Manchester Airport, followed by a number of urban areas that are known to concentrate high commercial activity.
|LSOA code||LSOA name||calls||x mean|
|E01018326||Cheshire Wes 034A||7846||8.7|
|35||Healthcare Practitioner Referral||16.2|
4. Analysis of Spatial Epidemiological Patterns
In this section, we investigate to what extent socio-economic and demographic factors, together with population related variables and data from location-based services (Foursquare) are associated with different types of ambulance incidents. We perform a correlation-driven analysis reporting pearson correlation scores between pairs of variables across the various information layers. We have applied the Bonferroni correction when measuring the statistical significance of all correlation measurements and they were all significant for the corresponding thresholds.
4.1. Incident frequencies and associations
The ten most frequent dispatch code call for each age group resulted in a combined list of ten dispatch codes as given in Table 3. They accounted for of all calls. Healthcare practitioner referral (Dispatch Code 35) is the most frequent and corresponds to calls made by doctors at healthcare facilities to transfer a patient to a hospital. In the analysis that follows next, we have filtered out calls that correspond to this incident code from the data as they relate primarily to operations within the health service (e.g. transfering patients to a hospital with more beds) and not to epidemiological traces we are interested in in the present work.
Figure 2 shows the pearson’s r correlation scores between the frequencies of all calls (Total Calls) per LSOA and the most frequent incident types (dispatch codes). A higher correlation score between two incident types implies a higher chance of them co-occuring spatially. breathing problems and chest pain correspond to one of the most related pairs (pearson’s ). As suggested by a correlation score of , overdose/poisoning related incidents are more likely to associate with the occurrence of psychiatric/suicide, convulsion/seizuers and unconscioous/fainting cases. breathing problems are more associated with chest pain incidents and sick person cases. These figures already indicate that different geographic areas may yield different epidemiological incidents in nature and hence the question becomes whether it is possible to identify the characteristics that makeup those areas and are contributing to these patterns.
4.2. Residential versus daytime populations
In answering the question what geographic and demographic features influence the number of calls to an ambulance service?, it seems reasonable to expect that, to some degree, the size of the population would be relevant. We present two population-based maps here. While LSOAs were created with a goal of their being equal in population size (and therefore, varying in area), there was some considerable variation as Figure 3 shows. In fact, population size ranged from 988 (an urban area in Southport) to 6137 (a largely rural area south of Lancaster but which contains the university campus of Lancaster University), with a median of 1520. The city centres of Manchester and Liverpool show one or two higher than average populations, but in general green colours which correspond to lower population areas dominate.
Populations fluctuate constantly as people move about their daily activities. A characteristic example of such process is commuting, where typically large populations move from the more peripheral areas towards urban centers. In the present context, we can hypothesize that ambulance calls are likely to happen in the areas where people are active and not simply in the areas where their residence is registered. With this in mind we have designed a new variable, namely daytime population which is the sum of the workplace population (described in the previous section) with residents younger than 16 and older than 74 year old. This aims to provide a proxy to the number of people active at a geographic area during working hours. Figure 3 shows the distribution of this feature geographically. Note that the colour scale is different from the previous plot showing residential population levels, with the range of daytime population being rather larger; from 132 in West Cumbria to 42357 in the City Centre of Manchester. Some areas of relatively low resident population have a much higher working population, with the reverse also being a possibility. Daytime and residential populations capture different temporal instances of population whereabouts and as we demonstrate in the following paragraphs are both important to estimating calls to the ambulance service.
The daytime population levels appears to be a much more important indicator to residential population levels as noted by the corresponding pearson’s r with the total number of calls ( for daytime population versus for residential). Daytime population is strongly associated with the frequency of unconscious/fainting incidents (), which are followed closely by traumatic injuries (), convulsions/seizures ) and falls (). In terms of how different age group resident numbers explain ambulance call variations, the age group 25-44 sticks out as the one that contributes consistently across different incident types (for most cases we have an ), with young kids of age 0-4 being associated to breathing problem related incidents () and the number of older people being the group mostly associated with falls (pearson’s ), in line with previous studies that have associated this age group with higher fall risk due to the presence of specific risk factors (e.g. weakness, unsteady gait, confusion and certain medications) (rubenstein2006falls).
4.3. Socio-economic indicators
We now investigate the relationship, between the index of multiple deprivation (IMD), the seven constituent metrics that it is comprised of and the frequencies and types of calls per LSOA. The pearson correlation scores between the different pairs of variables is shown in Figure 5. The Index of Multiple Deprivation (IMD) scores a pearson’s of with respect to breathing problems which is the most highly correlated incident type, followed by chest pain incidents and psychiatric/suicide incidents (). While the overall index provides a general notion of the deprivation levels of an area, its constituent metrics can shed light on the more specific factors that relate to ambulance calls. Income, employment, health and crime deprivation correlate highly with the incidents types noted above. Note that interestingly, the IMD score of an area does not yield an as high correlation score when considering other types of incidents such as falls, traumatic injuries or fainting. These results highlight how population in deprived areas are essentially more likely to suffer a serious medical condition, perhaps due to lack of access to preemptive care and lifestyle related factors. Links between deprivation of living standards have been identified before in the literature (bernard_emergency_1998; mccartney_how_2013; smith_relative_1992) with reported correlation score values in the range of across large geographic regions at national scale. Here we demonstrate that the link between deprivation and population health not only persists in smaller geographic scales, but in this case the geographic divide that exists amongst regions becomes larger.
4.4. Digital traces of human mobility
With the goal of understanding to what extent mobile web proxies of human urban activity such as Foursquare could capture geographic variations in ambulance calls we visualise the two sources of data in Manchester in Figure 6. For a set of LSOAs in the center of the city, we plot the predominant categories of calls and check-ins respectively. Distinctive geographical patterns of the types of incidents can be observed for both cities, with unconscious/fainting being the predominant epidemiological trend in the dense urban cores, whereas other types of emergencies are more characteristic of the peripheral areas. With comparison to human activity as derived from Foursquare check-ins, we can also note some activities which are more typical for the core as opposed to the periphery of the city such as nightlife and shopping.
We further quantify these relationships in Figure 7, where a correlation matrix of call types and check-in types is presented for the whole North West region whereas the ambulance service operates. We notice that specific types of human activity tend to be associated with particular calls on the small-scale of LSOAs. For example, the most common type of activity associated with the total number of calls is professional places. breathing problems tend to be called in areas with a high number of shopping and food check-ins, similar to chest pain. Cases of convulsions/seizures are mostly associated with food, nightlife, shops and work environments, as well as cases of falls, overdose/poisoning, traumatic injuries and reports of sick person. To a lesser extent, travel and outdoor activities were also related to such cases. However, the highest correlations found were between the food and nightlife categories and calls related to loss of consciousness (Pearson’s and respectively). Overall, most types activities correlate with the unconscious category of calls to a varying extent which was most expressed in dense urban centers where most calls are made (see Figure 1).
5. Predicting Ambulance Calls
Summary of linear regression results.
Our aim next becomes to predict the number of ambulance calls for each location by considering a set of the variables discussed in the previous section in an ordinary least squares (OLS) linear regression model. Formally, given an areaand then the number of calls that originated from the area, is set to be approximated by the linear following relationship:
represents a predictor variable, and whereis a vector of unknown parameters where is the number of input variables.
in this setting are unobserved scalar random variables (errors) which account for the discrepancy between the actually observed responses
and the predicted outcomes. To alleviate colinearity effects we have not used variables that are inherently related (e.g. age group frequencies and residential population). All variables have been standardized by substracting the corresponding mean and dividing by the standard deviation. As a metric of assessment for the prediction task we use the adjusted
which provides an indication of how much of the variance is explained by the model compared to the total variance of variabletaking into account the number of independent variables. Overall, we examine ten prediction tasks, one for the overall number of calls in an area, and one for each of the nine most popular incidents types.
In Table 4 we present the coefficients of the variables of the linear regression models built for four out of ten prediction tasks considered. The adjusted values are considerably high in most cases with an being achieved when the total number of calls is considered as the dependent variable. The model attains an for breathing problems and for cases of chest pain and sick person incidents the values of remained above . The lowest number was recorded for the cases of falls with an whereas the rest of the incident types were predicted with values above . In all cases, with the sole exception of breathing problems the Foursquare variable (check-ins) corresponded to a statistically significant case and in all such cases the sign of the variable was positive implying that a higher number of check-ins in an area is in general associated with a higher number of calls.
To assess the importance of the different information signals in explaining the variance of different types of incidents we run the following experiment. For each incident type, we removed each of the variables and measured the reduction in . To obtain then the relative importance of a variable we simply measured the reduction in associated to it with respect to the maximum reduction attained by any of the variables. The barplot in Figure 8 show the relative importance of each predictor for the total number of calls, whereas in Figure 9 we plot the results for all nine types of incidents. Notably, the index of multiple deprivation appears to be the most important variable in the majority of cases with daytime population explaining best falls and traumatic injuries cases. Foursquare check-ins correspond to the third most important variable when considering the total number of calls, whereas for incidents of unconscious/fainting it becomes the most important indicator. An explanation for the performance of the Foursquare variable is the fact that most unconscious/fainting incidents occur in the city centers and the service’s usage patterns tend to be associated with activity in commercial, food and nightlife areas. Its importance for the overdose/poisoning case of incidents where it scores higher even than daytime population activity points in this direction.
Our results highlight the opportunity that arises from using data from online media sources and the mobile web to power the operation of medical services critical for citizens. Limitations in using such data sources in the present context relate to biases in mobile application usage patterns amongst others. Daytime population levels have been an important predictor of ambulance calls and a clear improvement to simply using residential population information, yet it still represents a static signal about the activity levels of an area. Populations fluctuate constantly and so a promising future direction would be to exploit real time digital datasets from location-based services to model medical incident activity not only across geographies, but also over time. The importance of deprivation indicators in explaining geographic variations of ambulance calls provide an additional reminder of the large divides that exist in our society. Providing evidence through data driven analysis of population activity and government collected socio-economic indicators as we have done in the present work is an important step to bridge this gap by informing relevant policies.