Proceedings of the Spatial Information Research Centre’s 10th Colloquium 111 Connectionist Modelling of Asthma Incidence in New Zealand Abstract This paper describes an investigation of the patterns of self-reported asthma symptoms in relation to demographic and environmental factors in New Zealand. The subjects of the study consisted of 25,000 adults aged 20-44 who responded to a postal questionnaire. For each respondent, physical and social environmental conditions in the meshblock of residence were estimated using a Geographic Information System. The measured outcome was the 12-month prevalence of asthma. An artificial neural-network was constructed to model this outcome on the basis of the 13 environmental and demographic inputs, using a randomly selected sample of the data and tested on the remaining set of data. The modelling results indicated that there was different behaviour for the 20-25 and 26-44 age-groups, and separate neural-network models were constructed for each of these two age-groups. A set of inference rules were then extracted from each of the two neural networks. When applied to test data, the inference rules predicted the occurrence of asthma correctly in approximately 70% of the cases. Our approach may prove to be useful in simulating the effect of scenarios of environmental change on the occurrence of asthma. 1 Introduction A national survey carried out during 1991-3 found substantial regional variation in asthma prevalence among New Zealand general electorates (Lewis, 1997). Subsequent analyses found evidence of weak associations between asthma symptoms and environ- mental factors (Salmond, 1998; Hales, 1998). Hales et al. (1998), observed substantial differences in asthma prevalence between general electorates and across quartiles of physical environmental factors. There has been considerable debate about the possible role of various environmental factors in explaining temporal and geographical patterns in asthma prevalence (Devalia, Ruszniak et al. 1994; Martinez 1994; Seaton, Godden et al. 1994; Burr 1995; Newman-Taylor 1995; Strachan 1995; Balfe, Crane et al. 1996). Neural network models have been found to have better predictive accuracy than regression models in some data sets (Duh, 1998). In this study, we investigate the role of demographic, and environ- mental factors in patterns of adult asthma symptom prevalence in New Zealand by means of connectionist-based analysis. 2 Asthma prevalence in New Zealand Asthma prevalence data were obtained from an extensive survey of New Zealand citizens that was conducted during 1991-93. The European Commu- nity Respiratory Health Survey (ECRHS) measured adult asthma symptoms and severity in a number of countries, using standardised methods. New Zealand participated in the ECRHS, initially involving surveys in Auckland, Hawkes Bay, Wellington and Christchurch in 1991-2. The survey was subsequently extended to cover the whole country in 1993. The methodology for the survey has been described in detail elsewhere (Burney, Luczynska et al. 1994). Briefly, a one page questionnaire was mailed to 31,470 people aged 20-44, chosen from the 1991 New Zealand electoral roll, sampling at least 1 in 40 from each electorate. Addresses of registered voters in the Simon Hales * , Qingqing Zhou † , Simon Lewis ‡ , and Martin Purvis † ‡ Wellington School of Medicine, Wellington, New Zealand * Department of Public Health, Wellington, New Zealand † Information Science Department, University of Otago, Dunedin, New Zealand Presented at the 10th Colloquium of the Spatial Information Research Centre, University of Otago, New Zealand, 16-19 November, 1998