Francisco Sahli
Especialidad:
Physics-informed machine learning
Email:
fsahli1@uc.cl
Francisco es ingeniero civil mecánico y magíster en ciencias de la ingeniería de la Pontificia Universidad Católica de Chile. Actualmente se desempeña como profesor asistente del Departamento de Ingeniería Mecánica y Metalúrgica UC y del Instituto de Ingeniería Biológica y Médica de la misma casa de estudios.
PUBLICACIONES
Cardiac fibers play an essential role in the electrical and mechanical function of the heart, making them a critical parameter in cardiac modeling. However, identifying the 3D arrangement of the fibers in clinical practice still encounters some limitations. Solving the inverse problem of inferring the fiber orientations from electrophysiological data has been reported as a potential method for identifying patient-specific fiber orientations. In this work, we employ -Fibernet, a recently proposed Physics-Informed Neural Network (PINN) model, to reconstruct the fiber orientations on the right ventricle’s endocardial surface from a single activation map. The model equips the traditional PINN framework with an ensemble of parallel neural networks to cope with the uncertainty in the fiber approximations. Each ensemble member estimates the activation times and fiber orientations. The best fiber orientations are finally selected using a specific method to reduce the uncertainty of predictions. We evaluated the performance of the model in a simple propagation pattern and a more realistic propagation pattern incorporating the influence of the Purkinje network. Our results indicate the robustness of -Fibernet and its capacity to learn complex activation patterns and fiber distributions while trained with only a single activation map.
Structural optimization has the potential to save significant costs for buildings. However, most approaches require new algorithm implementations or a significant computational cost, which is often out of reach in terms of software and computational budget for structural engineers. This work addresses these issues through Bayesian optimization, where the structural analysis, design, and normative constraints are considered black-box functions. In this work, we optimize the shear-wall thicknesses of two finite-element building models in OpenSEES to reduce the total cost of concrete and steel reinforcement. We start by running a design of experiments phase to create initial datasets with different parameters. Then, we run the optimization by iteratively training Randomized Prior Networks as surrogate models for the cost and constraint and use Thompson sampling to estimate a batch of candidate optimal points at each iteration. Later, we propose multiple low-fidelity models to feed Bayesian optimization through multi-fidelity surrogate modeling to enhance the optimization performance. We analyze different combinations of dataset sizes and low-fidelity models, and we obtain lower costs when comparing single-fidelity methods with traditional metaheuristic methods. In addition, we found that multi-fidelity modeling can achieve similar or even better optimization results than high-fidelity-based optimization in a fraction of time. This work opens the door to increase the usage of structural optimization, leading to more efficient buildings.
Cardiac digital twins have shown promise to personalize treatments. However, there are multiple challenges to incorporate patient-specific information from non-invasive data. For instance, recovering the activation sequence in atria from the standard electrocardiogram (ECG) remains elusive. Recent studies have tackled this task on the ventricles, where the ECG signal is much stronger. This work presents a novel methodology to recover the atrial electrical activity with physics-informed neural networks. Instead of focusing on the activation times, we predict the direction of propagation of the electrical wave at each point with a neural network. Then, by solving a linear system for the Poisson equation, we recover the activation times that satisfy the anisotropic eikonal equation. The proposed methodology is compared with a methodology that predicts directly the electrical propagation and does not enforce the propagation model. We compare it to a traditional physics-informed neural network formulation, where the eikonal equation is only weakly imposed. We validate our methodology in a biatrial synthetic case using realistic lead fields for ECG calculation. We then learn the activation sequence from patient data, recovering a physiological activation pattern. We believe this is a first step toward digital twinning of the atria.
The identification of the Purkinje conduction system in the heart is a challenging task, yet essential for a correct definition of cardiac digital twins for precision cardiology. Here, we propose a probabilistic approach for identifying the Purkinje network from non-invasive clinical data such as the standard electrocardiogram (ECG). We use cardiac imaging to build an anatomically accurate model of the ventricles; we algorithmically generate a rule-based Purkinje network tailored to the anatomy; we simulate physiological electrocardiograms with a fast model; we identify the geometrical and electrical parameters of the Purkinje-ECG model with Bayesian optimization and approximate Bayesian computation. The proposed approach is inherently probabilistic and generates a population of plausible Purkinje networks, all fitting the ECG within a given tolerance. In this way, we can estimate the uncertainty of the parameters, thus providing reliable predictions. We test our methodology in physiological and pathological scenarios, showing that we are able to accurately recover the ECG with our model. We propagate the uncertainty in the Purkinje network parameters in a simulation of conduction system pacing therapy. Our methodology is a step forward in creation of digital twins from non-invasive data in precision medicine. An open source implementation can be found at http://github.com/fsahli/purkinje-learning.
Edificio de Innovación UC, Piso 2
Vicuña Mackenna 4860
Macul, Chile
Vicuña Mackenna 4860
Macul, Chile