CALIBRATION AND VALIDATION METHODS FOR AN AGENT-BASED SIMULATION TO MODEL CLOSTRIDIUM DIFFICILE TRANSMISSION DYNAMICS IN A HEALTHCARE INSTITUTION

Purpose: Agent-based simulation (ABS) is ideal for studying the transmission of Clostridium difficile infection (CDI) in a hospital, because it can model the interactions between patients, healthcare workers, and the hospital environment. The validity of parameters and accuracy of behavioral logic are particularly important for ABSs. However, complexities of disease transmission and scarcity of data can lead to difficulties in ascertaining the validity of ABSs. We develop an ABS to model the transmission and control of CDI in a healthcare institution, and we discuss methods used for calibration of unknown parameters and the validation of our ABS.

Method: We present an overview of our ABS and discuss methods we use to estimate unknown parameters, the calibration of a Markov model that describes natural disease progression of CDI in a patient, and the behavior of patients and healthcare workers. We use standard methods of validation and verification to ensure the development of an accurate model. We define an iterative calibration method where we use a brute-force search to estimate the transition probability matrix of the Markov model subject to constraints on long-run probabilities found from the literature. The resulting matrices are used in the simulation and scored using mean percentage error (MPE) of four statistics compared to benchmark values from published literature: colonization rate, CDI incidence, relapse CDI rate, and mortality rate. We impose additional constraints on the matrix values based previous calibration iterations, and then the calibration is rerun. Once sufficient convergence in the probabilities is attained, the calibration is halted.

Result: We run a scenario of commonly used strategies and average adherence rates. The resulting transition probability matrix yields model results with an average colonization incidence of 13.049% (MPE_{COLONIZATIONS}  = 0.127%), CDI incidence of 0.692% (MPE_CDI  = 1.308%), relapse CDI rate of 20.07% (MPE_RELAPSE = 1.082%), mortality rate of 4.36% (MPE_MORTALITY = 1.533%), and an average MPE of 1.01%.

Conclusion: Despite hurdles such as limited data, unknown parameters, and uncertainty, we can use computational methods to circumvent these issues and develop a sufficiently valid and calibrated simulation model of CDI transmission and control in a hospital.