—  Learn  —

Generation of in-silico virtual patient cohorts 

Background

The CircAdapt model of the human heart and circulation – developed by Maastricht University - enables fast and realistic simulation of cardiac mechanics and hemodynamics and is currently in use for medical education of cardiac (patho)physiology in many EU medical schools (www.circadapt.org). Multiple validation studies directly comparing simulated with animal experimental or clinical strain patterns have shown that CircAdapt accurately replicates myocardial deformation abnormalities caused by a wide variety of cardiac diseases – demonstrating its added clinical value and its ability to serve as a platform for realistic virtual patient simulations. Similarly, an advanced ultrasound simulation tool – developed at KU Leuven – enables to generate ultra-realistic echocardiography images from a finite element heart model by mimicking the image formation process in either 2D or 3D after spatiotemporal registration of the heart model with a real recording. This unique simulation tool is being used by the joint working group on standardization of speckle tracking echocardiography of the European Association on CardioVascular Imaging (EACVI) and the American Society of Echocardiography (ASE) as a follow-up of their previous activities on quality assurance of commercial tracking solutions based on synthetic ultrasound data.

Approach

Within MARCIUS, abnormal electrical activation will be simulated through the well-validated CircAdapt cardiovascular simulation framework (ESR1) by configuring different activation sequences of the ventricles. Similarly, myocardial pathologies such as regional or diffuse myocardial ischemia or scarring will be modelled by changing regional active and passive tissue properties. The effects of other more global patient-specific parameters such as heart rate, cardiac output, circulating blood volume and peripheral vascular resistance of the systemic circulation on regional myocardial deformation will also be included as degrees of freedom. Besides using these virtual strain traces for training of a classification neural network (cf. WP2), they will also be used to drive a kinematic heart model (ESR2) that can be fed into the ultrasound simulation pipeline of KUL. In this way, a unique database of cardiac disease substrates relevant to heart failure and their corresponding synthetic strain traces and ultrasound images will be generated. We plan to span a multi-dimensional parameter space with up to 500,000 simulations as input to supervised learning methods (cf. WP2). Non-deterministic effects and realistic levels of strain signal noise will be added to the simulated strain data to provide highly realistic myocardial strain data. Similarly, typical acoustic artefacts that impact strain analysis (e.g. reverberations, acoustic shadowing) will be integrated in ultra-realistic synthetic ultrasound recordings. Importantly, labelling of disease and annotation of the images is intrinsic to this learning approach given that the “ground truth” is known quantitatively for these virtual patients.

Outcomes & Synergies

The ability to generate virtual patient data with the underlying disease (at tissue-, organ- and system-scale) being known will be a “gold mine” for training machine learning algorithms, where data is king. We will be able to produce an immense database with detailed “ground truth” knowledge of underlying pathophysiology that spans both common and less common cardiac conditions. These ‘virtual patients’ will allow covering the feature space homogeneously, which is extremely hard to achieve in a clinical setting. Importantly, all data from ‘virtual patients’ will be complementary to true clinical recordings to avoid bias towards the simulation tool. In this way, biophysical models will guide the statistical methods; a trend recently seen in modern clinical trial design[1] and referred to in computer vision as ‘transfer learning’. The MARCIUS virtual patient simulation platform will provide a reliable and cheap alternative to animal experiments for basic research and training in the context of cardiovascular diseases. Hence, the proposed research will facilitate replacement, reduction and refinement (3Rs) of animal experiments. Further utilisation benefit will be achieved by integration of the virtual patient simulations into the intelligent cardiac ultrasound training products of MedaPhor, which will leverage practical cardiac ultrasound excellence in European medical centres.