Personalizing computational models to construct medical digital twins

Digital twin technology, pioneered for engineering applications, is being adapted to biomedicine and healthcare; however, several problems need to be solved in the process. One major problem is that of dynamically calibrating a computational model to an individual patient, using data collected from that patient over time. This kind of calibration is crucial for improving model-based forecasts and realizing personalized medicine. The underlying computational model often focuses on a particular part of human biology, combines different modeling paradigms at different scales, and is both stochastic and spatially heterogeneous. A commonly used modeling framework is that of an agent-based model, a computational model for simulating autonomous agents such as cells, which captures how system-level properties are affected by local interactions. There are no standard personalization methods that can be readily applied to such models. The key challenge for any such algorithm is to bridge the gap between the clinically measurable quantities (the macrostate) and the fine-grained data at different physiological scales which are required to run the model (the microstate). In this paper we develop an algorithm which applies a classic data assimilation technique, the ensemble Kalman filter, at the macrostate level. We then link the Kalman update at the macrostate level to an update at the microstate level that produces microstates which are not only compatible with desired macrostates but also highly likely with respect to model dynamics.