Data Mining Cancer Registries: Retrospective Surveillance of Small Area Time Trends in Cancer Incidence Using BaySTDetect

Li, Guangquan, Richardson, Sylvia, Fortunato, Lea, Ahmed, Ismail, Hansell, Anna, Toledano, Mireille and Best, Nicky (2011) Data Mining Cancer Registries: Retrospective Surveillance of Small Area Time Trends in Cancer Incidence Using BaySTDetect. In: 2011 IEEE 11th International Conference on Data Mining Workshops. IEEE, Piscataway, NJ, pp. 885-890. ISBN 978-1-4673-0005-6

Full text not available from this repository. (Request a copy)

Abstract

Space-time modelling of small area data is often used in epidemiology for mapping temporal trends in chronic disease rates. For rare diseases such as cancers, data are sparse, and a Bayesian hierarchical modelling approach is typically adopted in order to smooth the raw disease rates. Although there may be a general temporal trend which affect all areas similarly, abrupt changes may also occur in particular areas due to, for example, emergence of localized risk factor(s) or impact of a new health or screening policy. Detection of areas with "unusual'' temporal patterns is therefore important to flag-up areas warranting further investigations. In this paper, we present a novel area of application of a recently proposed detection method, Bays Detect, for short time series of small area data. Placed within the Bayesian model choice framework, Bays Detect detects unusual time trends based on comparison of two competing space-time models. The first model is a straightforward multiplicative decomposition of the area effect and the temporal effect, assuming one single temporal pattern across the whole study region. The second model estimates a local time trend, independently for each area. An area-specific model indicator is introduced to select which model offers a better description of the local data. Classification of an area local time trend as ``unusual'' or not is based on the posterior mean of this model indicator, which represents the probability that the common trend model is appropriate for that area. An important feature of the method is that the classification rule can be fine-tuned to control the false detection rate (FDR). Based on previous simulation results, we present some further insights of the model specification in relation to the detection performance in practice. Bays Detect is then applied to data on several different cancers collected by the Thames Cancer Registry in South East England to illustrate its potential in retrospective surveillance.

Item Type:	Book Section
Uncontrolled Keywords:	Bayesian spatio-temporal analysis, FDR, detection, disease surveillance
Subjects:	G100 Mathematics
Department:	Faculties > Engineering and Environment > Mathematics, Physics and Electrical Engineering
Depositing User:	Becky Skoyles
Date Deposited:	11 Mar 2014 12:39
Last Modified:	12 Oct 2019 22:29
URI:	http://nrl.northumbria.ac.uk/id/eprint/15706

Downloads