[Back]

G. Romstorfer, G. Schneckenreither:
"Using Open Source Geo-Data in Agent-Based Models of Health Care Utilization";
Talk: MATHMOD 2012 - 7th Vienna Conference on Mathematical Modelling, Wien; 2012-02-14 - 2012-02-17; in: "Preprints Mathmod 2012 Vienna - Full Paper Volume", F. Breitenecker, I. Troch (ed.); Argesim / Asim, 38 (2012), 421 - 422.

Introduction. The individual choice of a medical service provider depends on a variety of properties. Among
them are the type of service required, coverage of health insurance, quality of the facility, accumulation of multiple
treatment options and the distance to the patients domicile. The latter is an apparent yet fundamental property,
which - in a nation/system wide dynamic simulation of these decision processes - confronts us with different
problems. We discuss the integration of geographic information into an agent-based model for simulating the
dynamic distribution of patient-provider relations based on real health care and geographic data. Our work is
connected to a research cooperation of Vienna University of Technology and the Main Association of Austrian
Social Security Institutions.
Geographic Data. The basis of processing geographic information is a common coordinate reference system
(CRS) and tools like program libraries and database systems which are capable of performing geographic transformations,
calculations or queries. Open-source projects like PostGIS, OpenStreetMap (OSM) and QuantumGIS
have proven to be the best choice for our purposes. Not only do they provide a free of cost high quality state of
the art tool-chain for organising and processing geographic data but they also allow us to use a huge database of
cartographic data (OSM) and to combine data from different sources (e.g. Statistics Austria) and of different type
(e.g. raster data) into a single database.
The most difficult part of obtaining and preprocessing geographic information does clearly not concern cartographic
data and tools. The challenge lies in bringing health care data into a geographic context. Especially health
care data is very sensitive in terms of protection of privacy and thus must be anonymised, clustered or even blurred.
The very direct consequence of anonymised data is that we can no longer identify single individuals and their history
of visited providers (i.e. their decisions). This fact is unavoidable but does clearly not render any dynamic
simulation based on anonymised data useless in the first place [1].
A greater challenge are incomplete data sets, an inconsistent naming system or the complete lack of information
which would enable geographic allocation. The reason for these drawbacks are the great number of different
institutions involved in the processes of initial data acquisition and the lack of an immanent necessity to collect
geographic information. Furthermore we have to use reverse techniques to regenerate a reasonable distribution of
the locations of the patients domiciles. This can be achieved by randomly distributing the domiciles based on the
actual population density and the geographic information included in the health care data.
Techniques. For integrating geographic information into an agent-based model several steps are necessary:
Import of all required data into a single database. This involves decompression and extraction of data from
different sources and formats, alignment of nomenclature and transformation to a common CRS.
Basic geographic information (concerning providers and patients) is given in terms of addresses, postcodes or
assignments within a political structuring. Based thereon geographic locations (coordinates) must be assigned to
all agents of these groups by searching cartographic data (OSM) or by the technique mentioned at the end of the last
section. Latter is based on raster data on population count which must be combined with a political structuring. As
a consequence we have to solve a large number of point-in-polygon problems [2] in order to distribute the patients
in the resulting polygon grid (intersections of raster cells and political districts).
In order to speed up simulations, computationally intensive operations like finding the shortest route between
two locations or deciding whether a location lies within a certain area must be precalculated.
For integrating the resulting database into a dynamic simulation we have to use software libraries or modelling
frameworks which are capable of accessing database systems, provide basic geographic processing tools and allow
appropriate (even interactive) graphic display of results.
Steps one to three can be combined in an automated preprocessing tool, which allows to quickly react on the highly
dynamic geographic data which is available from OSM. For implementing a test model we rely on AnyLogic,
GeoTools and self-written Java classes.