Ensemble Kalman filter methods for agent-based medical digital twins

Abstract

A medical digital twin is a computational replica of some aspect of a patient’s biology relevant to patient health. It consists of a computational model that is calibrated to the patient and is dynamically updated using a stream of patient data. The underlying computational model is often, multi-state, stochastic, and combines different modeling platforms at different scales. Standard methods for data assimilation do not directly apply to such models. This is true in particular for ensemble Kalman filter (1, 2) methods, a common approach to such problems. This paper focuses on agent-based models (ABMs), a model type often used in biomedicine. The key challenge for any forecasting algorithm for this model type is to bridge the gap between (detailed) micro- and (summary) macrostates. This paper proposes a modified Kalman filter method to meet this challenge, providing a way to dynamically update the microstate of an ABM using patient measurements collected at the macro level.