Applied & Computational Mathematics seminar: Digital Twins for Clinical Applications: Complementarity of Mechanistic and Deep-Learning Modelling

Speaker:  Irene Vignon-Clementel, PhD

Title:  Digital Twins for Clinical Applications: Complementarity of Mechanistic and Deep-Learning Modelling

Abstract:   Digital twins are emerging as a promising framework for supporting personalized clinical decision-making by combining patient-specific data with computational models capable of describing, predicting, and interrogating physiological states. In this context, mechanistic and deep-learning approaches should not be viewed as competing paradigms, but rather as complementary components of a virtual human twin.

This seminar will illustrate this complementarity through several applications in cardiovascular and liver medicine. Mechanistic models, ranging from reduced-order whole-body circulation models to multi-scale flow and transport simulations, provide a physiologically interpretable representation of patient-specific systems and enable virtual interventions. Examples include the anticipation of postoperative portal hypertension following hepatectomy and the simulation-based design of shunts for pulmonary hypertension. However, personalization, uncertainty quantification, sensitivity and identifiability analyses, as well as inverse problems, may require a large number of computationally expensive model evaluations.

Deep-learning models can address this computational bottleneck by providing fast and differentiable approximations of mechanistic models. Neural-network surrogates are presented for reduced-order cardiovascular models, while physics- and discretization-informed graph neural networks are developed to accelerate multi-level myocardial perfusion simulations. In particular, embedding the finite-volume discretization directly into the learning process enables prediction across patient-specific geometries without relying on simulation-generated ground truth.

The seminar will further discuss multifidelity graph-based architectures that combine local and global information and incorporate physical constraints directly into the learning process. These approaches open new possibilities for efficient inverse modelling and data assimilation.

Together, these developments demonstrate how the integration of biophysical modelling, patient data and deep learning can bring computational digital twins closer to the speed and flexibility required for clinical decision support.

Time:  Friday, September 25, 10:00am-11:00am
Place:  Exploratory Hall, room 4106