Predictive Digital Twins for Health Monitoring: From Structural Safety to Personalized Medicine
Please login to view abstract download link
A digital twin (DT) can be defined as a personalized virtual representation of specific attributes of a natural or engineered system or process. It dynamically mirrors its physical counterpart by continuously assimilating sensor data and refining predictive capabilities accordingly. This continuous updating enables the simulation of what-if scenarios, supporting predictive decision-making tailored to maximize value. In this talk, we consider DTs modeled through probabilistic graphical models, specifically dynamic Bayesian networks, to encode the bidirectional interaction between the physical and virtual domains. DT adaptivity is introduced through a stochastic parametrization of the transition model, allowing for hierarchical online learning of state transition beliefs through Bayesian updates. Accordingly, we formulate parametric Markov decision processes and develop dynamic policies with precision updates. We present two DT applications. The first concerns the structural health monitoring, management, and maintenance planning of a railway bridge. Here, we use deep learning to assimilate vibration recordings for DT updating and reinforcement learning for sequential decision-making. Computational efficiency is ensured through an offline phase, where training data are generated using a reduced-order numerical model. The second application involves a patient-twin system for personalized clinical decision-making in congestive heart failure treatment. Medical covariates are continuously collected and assimilated with random forest classifiers to forecast future risks and guide the most appropriate treatment plan. The patient DT is calibrated and tested on experimental data from a cohort of patients in the UK Biobank.
