傳染病模擬模型在流行病學中被用來探索一些在現實中難以探究的問題。 其中,個體化模擬模型(Agent-based model)利用在電腦中的虛擬個體模擬由複雜行為組成的系統。 近年來,由於電腦運算技術的進步,個體化模擬模型有許多的應用產生,然而對於如何擬合與校正個體化模擬模型的研究甚少。 本研究利用連續時間貝氏網路(Continuous-time Bayesian Networks)發展了一組具有統計界面的傳染病個體化模擬模型,並進一步以過去的擬和架構為基礎,發展出一套擬合程序。 我們成功將遺傳演算法中的數值點突變(Numerical mutation)及參數分組策略(Blocking strategy)應用於序列蒙地卡羅法(Sequential Monte Carlo)中,使擬合程序可以處理大量參數且來源各異的資料。 最後,我們以肺結核的接觸者追蹤政策為例,使用易感受-感染者-復原者模型(Susceptible-Infectious-Recovery model)來演示我們為個體化模擬模型從模型建構、估計到預測所發展的實證架構。
The simulation models in epidemiology were developed to answer the questions which were not easy to solve by observational studies in the real world. In particular, Agent-based models (ABMs) were usually employed to deal with the complex system of disease transmission by simulating computational agents in the virtual world. However, the fitting scheme of ABMs is less developed than the applications.. With the aim of investigating disease dynamics and creating an interface for statistical analysis, we proposed a class of ABMs with Continuous-time Bayesian network, a temporal multivariate probability model. While retaining the strength of existing procedure for simulation model fitting based on sequential Monte Carlo, we set up an improved framework for fitting ABMs. We further synthesized the numerical mutation in genetic algorithm and the parameters augmentation in blocking Gibbs sampling in order to overcome the challenges of multidimensional parameters and multi-sources data. Using an example of Susceptible-Infectious-Recovery model for contact tracing in tuberculosis control, we briefly presented the properties of our proposed model and demonstrated its potential applications in the future. By including model construction, fitting, and forecasting, we formalized an empirical scheme for individual based models in simulating disease dynamics.