Predictive accuracy of particle filtering in dynamic models supporting outbreak projections

BMC Infectious Diseases

Research Article

Anahita Safarishahrbijari¹  Nathaniel D. Osgood¹  Aydin Teyhouee¹  Juxin Liu²  Cheryl Waldner³ 

[1] Department of Computer Science, University of Saskatchewan, 176 Thorvaldson Building, 110 Science Place, SK - S7N5C9, Saskatoon, Canada;Department of Mathematics and Statistics, University of Saskatchewan, College Drive, Saskatoon, Canada;Western College of Veterinary Medicine, University of Saskatchewan, Campus Drive, Saskatoon, Canada;

关键词: Particle filtering;    System dynamics;    Transmission model;    Compartmental model;    Stochastic;    Outbreaks;    Infectious diseases;    Communicable illness;    Empirical observations;   

DOI  :  10.1186/s12879-017-2726-9

 received in 2017-02-09, accepted in 2017-09-12,  发布年份 2017

来源: Springer

【 摘 要 】

BackgroundWhile a new generation of computational statistics algorithms and availability of data streams raises the potential for recurrently regrounding dynamic models with incoming observations, the effectiveness of such arrangements can be highly subject to specifics of the configuration (e.g., frequency of sampling and representation of behaviour change), and there has been little attempt to identify effective configurations.MethodsCombining dynamic models with particle filtering, we explored a solution focusing on creating quickly formulated models regrounded automatically and recurrently as new data becomes available. Given a latent underlying case count, we assumed that observed incident case counts followed a negative binomial distribution. In accordance with the condensation algorithm, each such observation led to updating of particle weights. We evaluated the effectiveness of various particle filtering configurations against each other and against an approach without particle filtering according to the accuracy of the model in predicting future prevalence, given data to a certain point and a norm-based discrepancy metric. We examined the effectiveness of particle filtering under varying times between observations, negative binomial dispersion parameters, and rates with which the contact rate could evolve.ResultsWe observed that more frequent observations of empirical data yielded super-linearly improved accuracy in model predictions. We further found that for the data studied here, the most favourable assumptions to make regarding the parameters associated with the negative binomial distribution and changes in contact rate were robust across observation frequency and the observation point in the outbreak.ConclusionCombining dynamic models with particle filtering can perform well in projecting future evolution of an outbreak. Most importantly, the remarkable improvements in predictive accuracy resulting from more frequent sampling suggest that investments to achieve efficient reporting mechanisms may be more than paid back by improved planning capacity. The robustness of the results on particle filter configuration in this case study suggests that it may be possible to formulate effective standard guidelines and regularized approaches for such techniques in particular epidemiological contexts. Most importantly, the work tentatively suggests potential for health decision makers to secure strong guidance when anticipating outbreak evolution for emerging infectious diseases by combining even very rough models with particle filtering method.

【 授权许可】

CC BY   
© The Author(s) 2017

【 图 表 】

40517_2023_273_Article_IEq9.gif

40517_2023_273_Article_IEq6.gif