HM Medical Clinic

 

Agile-online.org

Huerta, Schade, Granell (Eds): Connecting a Digital Europe through Location and Place. Proceedings of the AGILE'2014 International Conference on Geographic Information Science, Castellón, June, 3-6, 2014. ISBN: 978-90-816960-4-3 Analysing spatiotemporal patterns of antibiotics prescriptions Luise Hutka and Lars Bernard Technische Universität Dresden Professorship of Geoinformation Systems Dresden, Germany Abstract
The emergence of antibiotic resistances due to antibiotic residues in urban sewage systems is becoming an increasingly important issue. This paper presents a model for the spatiotemporal analysis of antibiotic inputs to derive spatiotemporal distribution patterns which are the basis for later predictions of future antibiotic inputs into the sewer system. To identify spatiotemporal distribution patterns of antibiotic prescriptions data statistical and GIS methods like time series and spatial cluster analysis are used. In order to find possible interrelationships the prescription data is combined with other influencing parameters (e.g. cases of respiratory infections) and tested for statistical correlations. Results show a pronounced seasonal course for three antibiotics of the macrolide group which also show high correlations with cases of respiratory infections in the study area. Further, results show that weekly data of respiratory infections by Google Flu Trends may be used as predictor variable to derive forecasts of future antibiotic inputs into the sewer system. Keywords: antibiotic prescriptions, spatiotemporal pattern recognition, drug residues, correlation analysis statistical tools and GIS. The paper starts in describing the study design to analyse spatiotemporal patterns of antibiotics Drug residues in sewage are an important issue in the context input into the sewage system using existing medical of the objectives of wastewater treatment [9, 12, 13]. In prescription data and further geodata. The remainder focuses general, it is assumed that an increased input of antibiotics on discussing the current results and on detecting appropriate into the environment promotes the formation of antibiotic- input variables – e.g. from crowd-sourced data – for a model resistant bacteria. If a pathogen is resistant to a particular to predict antibiotic release into the sewer system. antibiotic, taking this antibiotic in case of an infection with the pathogen is ineffective. In the end, the increasing antibiotic resistances together with the decreasing development of new Spatiotemporal Analysis of Antibiotic
antibiotics in recent years may lead to more and more Medications
antibiotics becoming ineffective, with the result that infectious diseases could spread again [19]. Thus, there are numerous Various studies demonstrate GIS to be a powerful tool for studies on the efficiency of various procedures regarding the analysing spatial and temporal distribution patterns of drug- specific behaviour of antibiotics and resultant antibiotic related health data. Cheng et al. [2] used local spatial resistance [1, 16, 17]. association statistics to examine the geographic variation of The project ANTI-Resist1 [5] researches the release of cardiovascular drug-prescribing patterns in Taiwan. Modarai antibiotics and potentially related appearance of antibiotic et al. [15] performed a local Moran's I analysis on annual resistances in the urban sewage system of the city of Dresden. opioid prescription sales aggregated by 3-digit zip codes and As a long term objective the project is meant to support the correlated this data with official data on opioid overdoses. As design of strategies to reduce the formation of antibiotic in the previous studies, also in the present work local resistances in urban wastewater. The project focuses on the Moran's I analysis is performed on prescription data, but in development of corresponding monitoring and warning applying a higher temporal and spatial resolution as the used systems. Within the project various aspects related to antibiotic prescription data are available at a weekly/monthly antibiotic fluxes and transports in urban wastewater systems resolution for 8 years and on an urban district level. A related are considered. First, the antibiotic prescriptions of medicines work on antibiotic prescriptions was published by Kern et al. are investigated. Second, models are being developed to [11] who studied the regional variation in outpatient antibiotic describe their release and transport within the sewage system use within Germany. In contrast to the work presented here, and third, the related emergence of antibiotic resistances Kern et al. analysed prescription data for only one year and on within the sewer and the water treatment facilities are studied the large-scale federal state level, further they did not apply using different measurement and observation methods. GIS-based spatial analysis. This paper describes the approach of spatiotemporal The ANTI-Resist project follows a twofold approach in analysis and modelling of antibiotic prescriptions using analysing spatiotemporal patterns of antibiotics emergence and related processes in urban wastewater: (1) Sewer measuring campaigns and related laboratory studies are

AGILE 2014 – Castellón, June 3-6, 2014 Figure 1: Schema of input model and prediction model conducted to better understand the actual antibiotic fluxes and Input Data
the genesis of antibiotic resistance within the sewer system and (2) data driven (statistical) models are getting designed to Basis for the input model are data about the ambulant serve as best estimated guesses on where antibiotics are being antibiotic prescriptions that have been provided by one of the released and how they are transported through the sewage German compulsory health insurances (AOK PLUS). These system. Figure 1 sketches the conceptual frame for these data are available in a weekly and monthly temporal models and consists of two components: resolution for the period from 2005 to 2012. The data are An input model has been designed to analyse provided for the 64 urban districts of the city of Dresden and historical prescription data, socio-economic data aggregated into three age groups: 0 to 14 years, 15 to 64 and environmental data to identify typical years, 65 and more years. The following antibiotic substances of various active ingredient groups are examined: amoxicillin, medications, to estimate related antibiotic fluxes azithromycin, cefuroxime, ciprofloxacin, clarithromycin, into to the sewer system and to evaluate these levofloxacin-ofloxacin, estimates against different measurements within the The input model then serves as the basis for a Several issues arose when analysing the prescription data prediction model to estimate the future release and from the health insurance. First, the supplied data comprise fluxes of antibiotics into the sewer system. The only the patients of this single insurance company, which prediction model shall serve to alert the urban waste represent about 41% of the population of Dresden. So the water treatment and environmental agencies about prescription data is extrapolated to the total population of the potential occurrence of antibiotic peaks in the Dresden. Second, due to medical data protection issues sewage and the released waste water. As the prescription amounts between 1 and 3 are anonymised. In the treatment of antibiotics is not part of the operational calculations this problem is handled by setting all anonymised waste water treatment these alerts could also trigger values to the minimum amount of 1. Third, the data can only related specific measurements in the sewage plant. be delivered with a delay of at least one year. Consequently The main objective of the input model is to use spatial up-to-date official prescription data are not available for the statistical tools to investigate the variation of antibiotic inputs studies. This fact is seriously hampering the prediction of in time and within the urban districts of Dresden as well as future antibiotics release into the urban sewer system. To correlate it with other influencing factors that might have an overcome the latter problem it was necessary to identify data impact at the variation of the antibiotic prescriptions. This of other influencing factors that are related to antibiotic paper focuses the design and results of the input model and prescriptions and would eventually serve as a proxy for a will discuss some initial design ideas for the prediction model. prediction of antibiotic medication. The identified major influence factors are presented in the following.

AGILE 2014 – Castellón, June 3-6, 2014 Several studies have shown that there is a correlation time series into three components: seasonal variation, long- between respiratory infections and antibiotic prescriptions. On term trend and random noise (remainder component) [3]. the one hand for the majority of treated cases of respiratory Thus, the temporal variation of the prescriptions of the diseases antibiotics are prescribed. This is evident especially individual antibiotic substances in the overall city as well as in during the annual cold waves and winter flu season, which the individual urban districts could be examined as to whether regularly are accompanied by a significant increase of there are certain long-term trends and periodic seasonal antibiotic prescriptions [4, 8, 20]. On the other hand often a patterns. An example of the result of such a time series secondary bacterial infection follows a flu infection, as the analysis is shown in Figure 2. organism is already weakened due to the fight against the viruses. Therefore bacteria can more easily lead to further Figure 2: An example for the resulting components of a time infections that are then often treated by the use of antibiotics – series analysis in R using the STL function (sum of all antibiotic substances, 2008 – 2010) prophylactically, without bacterial caused symptoms being already present [10, 18]. For these reasons contemporary available representative data about current cases of respiratory diseases have been considered as potential proxies for the input model. Such data is provided by the flu trends portal from Google2. Ginsberg et al. [6] have analysed billions of Google search requests and determined that there is a high correlation between the frequency of particular search terms and the actual number of patients with influenza-like symptoms at a time. For the validation of their results they used historical data of traditional influenza surveillance systems. For Germany Google Flu Trends provides the weekly cases of respiratory infections per 100,000 inhabitants at a Federal State level from 2003 up to the current week [7]. Other influencing factors that have been identified and incorporated into the input model are meteorological (temperature, precipitation, etc.) and sociodemographic (population, employment structure) parameters. For the meteorological parameters the Regional Climate Information System for Saxony, Saxony-Anhalt and Thuringia (ReKIS)3 is used. This database contains – for the city of Dresden – data for 3 climate stations and 3 precipitation stations in a daily resolution for the years 1961 to 2013. At the most comprehensive station about 16 parameters are determined. For the sociodemographic parameters data from the Dresden statistics office is used4. It provides yearly data at an urban district level about the population, discriminated by sex and nine age groups and data on employees and unemployed persons as relative proportions of the relevant age group. The health insurance company AOK PLUS provided data on the age structure of their patients in 5 age groups at the urban ArcGIS (ESRI ArcGIS 10.2) with its geostatistical methods district level for the year 2011. Most of the input data for pattern recognition has been used to analyse the spatial described here can be explored via the ANTI-Resist distribution of the antibiotics prescriptions within the Dresden urban districts. ArcGIS offers several functions for the analysis of local spatial patterns: Hot Spot Analysis (Getis- Ord Gi*)6 and Cluster and Outlier Analysis (Anselin Local Morans I)7. Figure 3 and 4 show examples of the results of the respective cluster method. The Hot Spot Analysis results in In a first step the antibiotic prescriptions have been modelled statistically significant clusters of similarly high values regarding their temporal and spatial distribution using various (Figure 3 - red) or low values (Figure 3 - blue) for different statistical and GIS-based methods. confidence levels. In contrast, the result of the Cluster and The temporal analysis has been carried out by time series Outlier Analysis shows not only statistically significant analysis in R using the STL function, which decomposes a clusters of high values (Figure 4 - red) or low values

AGILE 2014 – Castellón, June 3-6, 2014 (Figure 4 - blue) but also districts of high values surrounded by low values (Figure 4 - yellow) and vice versa (Figure 4 - white). As the Cluster and Outlier Analysis seemed more Time series analysis
meaningful and as it also provides hints to outlier districts, this method has been selected. The individual antibiotic substances are prescribed differently In this way, the spatial and temporal prescribing patterns of throughout the year, depending on the application or bacteria antibiotics in Dresden and their changes over time have been causing certain infections. The selection of a suitable determined. In a next step possible explanations for the antibiotic is at the discretion of the attending physician. This occurrence of specific distribution patterns should be is also reflected in the temporal analysis of the antibiotic identified. Therefore the antibiotic prescription data are prescriptions data for the years 2005 to 2011, which showed analysed to search for statistical correlations (next section) that the temporal prescription behaviour, especially the considering the above mentioned influencing factors. The seasonal course, differs according to the individual antibiotic correlation analyses are mostly performed via simple linear substances. While some substances show a more or less Pearson correlation (using SPSS Statistics 21). pronounced seasonal pattern (amoxicillin, azithromycin, The input model still needs to be validated with actual clarithromycin, doxycycline, roxithromycin), others are measurements in the sewer system made within the project. prescribed relatively evenly over the year (ciprofloxacin, However, a successful validation study strongly depends on clindamycin) or have no specific pattern (cefuroxime, the identification of appropriate measuring points that can be levofloxacin-ofloxacin, defined as being comparable with the prescription data at the sulfamethoxazole-trimethoprim). The clearest seasonal pattern urban district level. is shown by the antibiotics of the macrolide group (azithromycin, clarithromycin, roxithromycin) with maximum Figure 3: Hot Spot Analysis of azithromycin prescriptions in prescriptions in the winter season and minimum values during the summer months. Regarding the trend component, most antibiotics indicate a decreasing trend. Only amoxicillin, cefuroxime and levofloxacin-ofloxacin reveal an increasing trend, while for azithromycin and clindamycin there is almost no trend apparent. Cluster and Outlier analysis
The result of the Cluster and Outlier analysis in ArcGIS depends on the selected distance and spatial relationship of the neighbouring features by which the algorithm calculates the clusters. Therefore, the analysis was first carried out with different parameter settings to find the appropriate preferences. conceptualization ZONE_OF_INDIFFERENCE is chosen, where a threshold value specifies whether to include or exclude neighbours and Figure 4: Cluster and Outlier Analysis of azithromycin where neighbours are weighted by the Inverse Distance prescriptions in January 2009 conceptualization of spatial relationship seemed most suitable for the analysis of the antibiotic prescriptions within the Dresden urban districts, as the polygons of the urban districts have different sizes and the (daily) mobility of citizens may lead to movements across several district boundaries. To cope with these cases it is recommended to use a distance-based conceptualization with a smooth transition as by ZONE_OF_ INDIFFERENCE. The ArcGIS tool Incremental Spatial Autocorrelation8 is used for choosing an appropriate threshold distance. The tool calculates the spatial autocorrelation for various distances, to determine the distance at which the spatial processes (the clustering) are most pronounced. This resulted in proper threshold distances of 2800 and 6400 meters for the data at hand, depending on the scale considered for the resulting clusters. AGILE 2014 – Castellón, June 3-6, 2014 Figure 5: Weekly cases of respiratory infections and macrolide prescriptions 2009 – 2011 Source: respiratory infections data: Google Flu Trend; prescription data: AOK PLUS (AOK Sachsen und Thüringen) The outcome of these analyses provided resulting cluster Table 1: Pearson correlation coefficients (R) for monthly maps, which are very different according to the considered Google Flu Trends data and substance-specific antibiotic substance and the given point in time. There were antibiotic prescriptions from 2005 to 2011 (n = 84) hardly any general patterns for all data. Therefore, the active ingredient individual cases have to be considered. Nonetheless, there are some urban districts, mostly in the central region of Dresden, that are more often part of a significant cluster than others (see also Figure 4). Thus, from a sewage treatment perspective, these cluster areas are the neighbourhoods that should be given priority and be investigated in more detail. Levofloxacin-ofloxacin Correlation analysis
As stated above data about respiratory infections are Phenoxymethyl penicillin considered as a proxy for antibiotic prescriptions. A correlation analysis on the monthly Google Flu Trends data and the prescriptions for all individual antibiotic substances has been performed (Table 1). As result there are three * The correlation is significant at the 0.05 level. substances presenting a significant strong correlation with the ** The correlation is significant at the 0.01 level. cases of respiratory infections: azithromycin (r = 0.94, p < 0.01), roxithromycin (r = 0.83, p < 0.01) and clarithromycin (r = 0.76, p < 0.01), all belonging to the Conclusion and Outlook
macrolide group. These three antibiotics show a seasonal pattern similar to the annual wave of influenza. The presented study succeeded in identifying spatial and Consequently the correlation analysis for the three temporal patterns of antibiotic descriptions, offering a macrolide substances has been repeated with higher resoluted promising path for future predictions of antibiotic releases in weekly data to include a wider sample and to create a basis for urban waste water. This work also demonstrated a statistical prospective weekly predictions (Figure 5). That way the significant correlation between respiratory infections and correlation coefficient has been improved for roxithromycin (r prescriptions of antibiotic substances of the macrolide group. = 0.88, p < 0.01) and clarithromycin (r = 0.90, p < 0.01), for However, it should be noted that it is not known when exactly azithromycin (r = 0.91, p < 0.01) it remains almost as high as a respiratory infection occurs within a week, how long it lasts with the monthly data. and when exactly any antibiotics are prescribed or a secondary infection occurs. Moreover, as a recent study pointed to some issues in the usage of Google Flu Trends [14], further AGILE 2014 – Castellón, June 3-6, 2014 investigation are required to validate the first results, epidemics using search engine query data. Nature, presented in this paper. Additional correlation analyses 457(7232): 1012-1014, 2009. considering other influencing factors as meteorological and sociodemographic parameters are ongoing. [7] Google, editor. Google Flu Trends. Frequently asked Future work will focus (1) on the validation of the input model using the results of the ANTI-Resist measurement campaigns and (2) on the development of a prediction model last accessed 02/2014. to derive forecasts of the expected antibiotics input into the sewer system. Predictions will be deduced by the combination [8] H. Goossens, M. Ferech, R. Vander Stichele and M. of the findings of the input model with up-to-date information Elseviers. Outpatient antibiotic use in Europe and using regression and interpolation functions. Based on the association with resistance: a cross-national database shown high correlation between respiratory infections and study. The Lancet, 365(9459): 579-587, 2005. macrolide prescriptions, the weekly available Google Flu Trends data can be integrated into a simple linear regression [9] T. Heberer. Tracking persistent pharmaceutical residues model to derive prediction intervals of expected antibiotic from municipal sewage to drinking water. Journal of inputs into the sewer system for the same week. Thus, the Hydrology, 266(3): 175-189, 2002. final results could be useful for the sewage treatment plant, which could initiate prompt provisions to increasing [10] S. Herold. Pathogenese, Klinik und Therapie der antibiotics input events, as well as for the public health sector, which could regulate the prescription behaviour of antibiotics Intensivstation. Pharmazie in unserer Zeit, 40(2): 115- in the appropriate way. 119, 2011. (In english: Pathogenesis, clinic and therapy [11] W.V. Kern, K. Nink, M. Steib-Bauert and H. Schröder. Regional variation in outpatient antibiotic prescribing in Germany. Infection, 34(5): 269-273, 2006. The project is funded by the Federal Ministry for Education and Research (BMBF) and is part of the program "Research [12] K. Kümmerer. Antibiotics in the aquatic environment – a for Sustainable Development". The fruitful cooperation with review – part I. Chemosphere, 75(4): 417-434, 2009. our ANTI-Resist project partners is gratefully acknowledged. Special thanks go to Daniel Kadner for developing the ANTI- [13] K. Kümmerer. Antibiotics in the aquatic environment – a Resist Geoportal. review – part II. Chemosphere, 75(4): 435-441, 2009. References
[14] D. Lazer, R. Kennedy, G. King and A. Vespignani. The Parable of Google Flu: Traps in Big Data Analysis. [1] A. L. Batt, S. Kim and D. S. Aga. Comparison of the Science, 343 (6176): 1203-1205, 2014. occurrence of antibiotics in four full-scale wastewater treatment plants with varying designs and operations. [15] F. Modarai, K. Mack, P. Hicks, S. Benoit, S. Park, C. Chemosphere, 68(3): 428-435, 2007. Jones, S. Proescholdbell, A. Ising and L. Paulozzi. Relationship of opioid prescription sales and overdoses, [2] C. L. Cheng, Y. C. Chen, T. M. Liu and Y. H. K. Yang. North Carolina. Drug and alcohol dependence, 132(1): Using spatial analysis to demonstrate the heterogeneity of the cardiovascular drug-prescribing pattern in Taiwan. BMC public health, 11(1): 380, 2011. [16] B. Pauwels and W. Verstraete. The treatment of hospital wastewater: an appraisal. J Water Health, 4: 405-416, [3] R. B. Cleveland, W. S. Cleveland, J. E. McRae and I. Terpenning. STL: A seasonal-trend decomposition procedure based on loess. Journal of Official Statistics, [17] G. Qiu, Y. Song, P. Zeng, L. Duan, and S. Xiao. 6(1): 3-73, 1990. Combination of upflow anaerobic sludge blanket (UASB) and membrane bioreactor (MBR) for berberine [4] P. Davey, C. Pagliari and A. Hayes. The patient's role in reduction from wastewater and the effects of berberine the spread and control of bacterial resistance to on bacterial community dynamics. Journal of hazardous antibiotics. Clinical Microbiology and Infection, 8(s2), materials, 246: 34-43, 2013. [18] T. Schaber. Diagnostik, Therapie und Prävention der [5] Forschungsverbund Public Health Sachsen und Sachsen- Influenza (Virusgrippe). Pneumologie, 57(01): 27-33, Anhalt. ANTI-Resist - Online available: 2003. (In english: Diagnosis, Therapy and Prevention of last accessed 02/2014. [6] J. Ginsberg, M. H. Mohebbi, R. S. Patel, L. Brammer, [19] B. Spellberg, R. Guidos, D. Gilbert, J. Bradley, H. W. M.S. Smolinskiand L. Brilliant. Detecting influenza Boucher, W. M. Scheld, J. G. Barlett and J. Edwards. AGILE 2014 – Castellón, June 3-6, 2014 The epidemic of antibiotic-resistant infections: a call to action for the medical community from the Infectious Diseases Society of America. Clinical Infectious Diseases, 46(2): 155-164, 2008. [20] S. Stone, R. Gonzales, J. Maselli and S. R. Lowenstein. Antibiotic prescribing for patients with colds, upper respiratory tract infections, and bronchitis: a national study of hospital-based emergency departments. Annals of emergency medicine, 36(4): 320-327, 2000.

Source: https://agile-online.org/Conference_Paper/cds/agile_2014/agile2014_111.pdf