[Paper Review] Foursquare to The Rescue: Predicting Ambulance Calls Across Geographies

Anastasios Noulas, Colin Moffatt+2 more|arXiv (Cornell University)|Jan 29, 2018

Data-Driven Disease Surveillance24 references3 citations

TL;DR

This study proposes a multi-source predictive model for ambulance call volumes across Lower Super Output Areas (LSOAs) in North West England, integrating Foursquare check-ins, daytime population estimates, and deprivation indices. It demonstrates that Foursquare check-ins significantly improve predictions—especially for nighttime incidents like unconscious/fainting and overdose/poisoning—outperforming traditional residential population data.

ABSTRACT

Understanding how ambulance incidents are spatially distributed can shed light to the epidemiological dynamics of geographic areas and inform healthcare policy design. Here we analyze a longitudinal dataset of more than four million ambulance calls across a region of twelve million residents in the North West of England. With the aim to explain geographic variations in ambulance call frequencies, we employ a wide range of data layers including open government datasets describing population demographics and socio-economic characteristics, as well as geographic activity in online services such as Foursquare. Working at a fine level of spatial granularity we demonstrate that daytime population levels and the deprivation status of an area are the most important variables when it comes to predicting the volume of ambulance calls at an area. Foursquare check-ins on the other hand complement these government sourced indicators, offering a novel view to population nightlife and commercial activity locally. We demonstrate how check-in activity can provide an edge when predicting certain types of emergency incidents in a multi-variate regression model.

Motivation & Objective

• To improve spatial prediction of ambulance call volumes at fine geographic granularity (LSOAs) in North West England.
• To investigate the role of non-traditional data sources—particularly Foursquare check-ins—in explaining geographic variation in emergency medical demand.
• To assess the relative importance of demographic, socio-economic, and real-time mobility indicators in predicting different types of ambulance incidents.
• To evaluate whether dynamic population activity (daytime population) outperforms static residential population in predicting emergency service demand.
• To inform public health policy and emergency service planning through data-driven insights into spatial disparities in medical demand.

Proposed method

• Constructed a multi-variate linear regression model: $ y_i = extbf{x}_i^T oldsymbol{eta} + oldsymbol{ u}_i $, where $ y_i $ is the number of ambulance calls in area $ i $, and $ extbf{x}_i $ includes predictors like daytime population, deprivation index, and Foursquare check-ins.
• Standardized all predictor variables by subtracting the mean and dividing by the standard deviation to reduce multicollinearity.
• Used adjusted $ R^2 $ as the primary evaluation metric to assess model performance across ten prediction tasks: one for total calls and nine for specific incident types.
• Performed variable importance analysis by sequentially removing each predictor and measuring the resulting drop in $ R^2 $, normalizing by the maximum drop to compute relative importance.
• Integrated open government datasets (e.g., deprivation indices, population demographics) with Foursquare’s location-based check-in data to capture real-time urban activity patterns.
• Defined daytime population as the sum of residents under 16, workers, and over 65 in each LSOA, using this as a proxy for non-residential activity.

Experimental results

Research questions

• RQ1 How do daytime population levels compare to residential population in predicting ambulance call volumes across LSOAs?
• RQ2 To what extent do Foursquare check-in patterns improve the prediction of specific ambulance incident types, especially those linked to nightlife or commercial activity?
• RQ3 Which socio-economic and demographic factors—particularly the Index of Multiple Deprivation—most strongly correlate with higher ambulance call volumes?
• RQ4 How do different types of emergency incidents (e.g., falls, unconscious/fainting, overdose) co-occur spatially, and what predictors best explain their geographic distribution?
• RQ5 Can real-time digital mobility data from location-based services like Foursquare serve as a reliable proxy for dynamic population activity in emergency demand modeling?

Key findings

• The model achieved an adjusted $ R^2 $ of 0.702 when predicting total ambulance calls, indicating strong explanatory power across all input variables.
• Daytime population levels showed a Pearson correlation of $ r = 0.68 $ with ambulance call volumes, significantly outperforming residential population ($ r = 0.18 $).
• Foursquare check-ins were a statistically significant predictor for 8 out of 9 incident types, with a positive coefficient indicating higher check-in frequency correlates with higher call volume.
• For unconscious/fainting incidents, Foursquare check-ins were the most important predictor, outperforming both daytime population and deprivation index.
• The Index of Multiple Deprivation was the most important variable for the majority of incident types, especially for breathing problems, chest pain, and psychiatric/suicide-related calls.
• Foursquare check-ins improved prediction accuracy for overdose/poisoning incidents more than daytime population, highlighting their value in capturing nightlife and commercial activity patterns.