Effective use of evolutionary computation to parameterise an epidemiological model

Abstract

Predictive epidemiological models are able to be used most effectively when they have first been shown to fit historical data. Finding the right parameters settings for a model is complex: the system is likely to be noisy, the data points may be sparse, and there may be many inter-related parameters. We apply computational intelligence and data mining techniques in novel ways to investigate this significant problem. We construct an original computational model of human papilloma virus and cervical intraepithelial neoplasia with the ultimate aim of predicting the outcomes of varying control techniques (e.g. vaccination, screening, treatment, quarantine). Two computational intelligence techniques (genetic algorithms and particle swarm optimisation) are used over one- stage and two-stage optimisations for eight real-valued model parameters. Rigorous comparison over a variety of quantitative measures demonstrates the explorative nature of the genetic algorithm (useful in this parameter space to support the modeller). Correlations between parameters are drawn out that might otherwise be missed. Clustering highlights the uniformity of the best genetic algorithm results. Prediction of gender-neutral vaccination with the tuned model suggests elimination of the virus across vaccinated and cross-protected strains, supporting recent Scottish government policy. This preliminary study lays the foundation for more widespread use of computational intelligence techniques in epidemiological modelling.