47 min
Seminar, Hosted by Stanford MedAI
Using physics informed neural network methods to build digital twins of cardiac hemodynamics
UC Berkeley and UCSF
Recording
Abstract
Digital twins in medicine allow in-silico simulations that mimic a patient's condition and counterfactual outcomes under hypothetical interventions. If they can be fine-tuned to individual patients from non-invasive data, they could remove the need for invasive procedures or uncertain treatment decisions. This talk demonstrates a method that identifies patient-specific digital twin parameters using deep learning and non-invasive health data. It treats the digital twin as a composite inverse problem whose structure resembles pretraining and fine-tuning in self-supervised learning, and presents a physics-informed self-supervised algorithm that pretrains a neural network to act as a differentiable simulator of cardiac hemodynamics. A model is then trained to reconstruct physiological measurements from echocardiogram videos while constrained by the physical equations learned in pretraining. The work belongs to a broader literature on machine learning methods that learn interpretable latent states by infusing domain physical knowledge into data-driven models.