Predicting the risk of cerebrovascular thrombosis in patients with heart disease



This was the main goal of a study published in Frontiers in Physiology by a team of researchers led by Oscar Camara, a member of the PhySense group, attached to the BCN MedTech research unit, of which Xabier Morales is the first author, with international participation from academic and research institutions from Denmark and France.

Due to the general ageing of the world population, cardiovascular diseases have become a major cause of death. The most common heart diseases include atrial fibrillation, a type of arrhythmia with a cumulative lifetime development risk above 30% in individuals of European ancestry.

Atrial fibrillation is characterized by chaotic electric activity, which leads to irregular contraction and wall rigidity of the left atrium, preventing effective flow of the blood to the ventricles. These changes in haemodynamics, among other factors, favour the formation of blood clots or thrombi, which exposes atrial fibrillation patients to a greater risk of suffering cerebrovascular accidents.

Surprisingly, up to 99% of all ischemic strokes of cardiac origin in patients with atrial fibrillation form at the left atrial appendage (LAA), a heterogeneous, tubular structure derived from the anterior wall of the left atrium. It has been hypothesized that the specific morphology of each patient’s appendage could be the main catalyst of thrombogenesis by promoting blood stagnation. Numerous studies have attempted to categorize appendages in terms of their morphology and relate each group with its own risk of thrombosis, but the results have proved ambiguous, since the morphological guidelines for the morphological classification of appendages are often completely subjective, stressing the need for more systematic, observer-independent analytical procedures.

Recently, the adoption of computational fluid dynamics (CFD) has contributed great knowledge about the interaction between morphology, haemodynamics and LAA thrombosis. CFD solves the physical equations that describe the behaviour of fluids, in this case blood, to find out its speed and direction throughout the cardiac cycle. However, although these CFD simulations provide a high degree of accuracy, they are notoriously slow, require a huge amount of computational resources and involve a lot of pre-processing, resulting in studies with very small samples.

Geometric deep learning can be especially useful in computational cardiology as you can work directly on the geometric meshes and it avoids the need for correspondence between patients

In response to these limitations, deep learning has begun to be used as an effective substitute for complex physical systems such as fluid dynamics. Deep learning involves overlapping several layers of artificial neurons that function like small computers that can alter the input value they receive through a simple operation. Although these models require enormous amounts of data to be accurate, once trained they can perform inference almost instantly.

"Hence neural networks fit our problem perfectly as we are seeking to quickly assess the risk a patient has of a thrombus without having to wait to complete the simulation. Once trained, we will input to the network the data on the morphology of each patient’s appendage, and the network will be able to predict thrombotic risk parameters we would obtain with the simulations instantly", says Xabier Morales , first author of a study recently published in Frontiers in Physiology. "Further still, during training, the network will learn by itself the most relevant morphological characteristics in respect of the risk of thrombosis, thus avoiding defining ourselves the morphological characteristics of interest completely subjectively and impartially", adds the member of the PhySense, research group of the BCN MedTech Unit at the UPF Department of Information and Communication Technologies ( DTIC ). The study has involved international participation by academic and research centres in Denmark and France.

Triangular meshes to represent anatomical images

However, one of the main bottlenecks that hinder the widespread use of these models is that conventional neural networks are not well suited to data that are not structured in regular grids such as images. The most efficient means of representing 3D objects such as the anatomy of human organs is through triangular meshes. Unlike the pixels of an image, in a geometric mesh the number of nodes adjacent to a vertex may vary. An equivalent example would be a list of the followers of each Twitter user, which will obviously vary for each person, while a pixel in a 2D image will always have eight neighbours regardless of its position. This greatly complicates defining operations like convolutions, which are the cornerstone of deep learning on the meshes themselves.

Generally speaking, conventional deep learning methods require generating correspondence through 2D mapping or regularizing the structure in a similar way. In this regard, geometric deep learning includes a whole set of recent methods that allow extending these operations to complex structuring data, which allows working directly on the meshes without any prior processing.

Neural networks can be effective substitutes for simulations of complex dynamic systems obtaining instant predictions, opening the way to future, real-time applications

Therefore, " in this study we seek to develop a deep learning framework capable of predicting the endothelial cell activation potential (ECAP), an in-silico index , typically derived from CFD simulations, solely from the patient-specific LAA morphology. To this end, we compare two conventional approaches to deep learning, "fully-connected" networks (all neurons are connected to all neurons of the next layer) and convolutional neural networks, which require prior processing and transformation of the meshes to regularize their structure, with a geometric deep learning architecture that can work directly on the native structure of the triangular mesh", points out project leader Oscar Camara.

The geometric model showed greater accuracy

To train networks, the researchers completed a total of 370 CFD simulations consisting of two datasets. An initial simpler dataset, using an artificial atrium and mixing real and synthetic appendages derived from statistical models, and a second, more complete dataset of case studies that considered the patient’s entire atrium.

First, the three models were able to predict ECAP instantly solely from the anatomical mesh of the appendage in the simplified dataset, effectively omitting the need to run computational fluid dynamics simulations. However, the geometric model showed great accuracy and significantly higher capacity for generalization. As it was possible to work directly on the native form of data, the geometric network proved to be able to learn more universal morphological characteristics as it obtained similar results to those obtained on real appendages training only on synthetic data. Moreover, this same model exhibited good predictive capability even in more advanced simulations with improved boundary conditions and including the entire LA anatomy.

These results could lay the foundation for real-time monitoring of LAA thrombosis risk in the future and open exciting avenues for future preoperative applications and interfaces in which a clinical user could interactively change settings of a left atrial appendage occluder device and instantaneously assess the associated risk.

Reference work:

Xabier Morales Ferez, Jordi Mill, Kristine Aavild Juhl, Cesar Acebes, Xavier Iriart, Benoit Legghe, Hubert Cochet, Ole De Backer, Rasmus R. Paulsen, Oscar Camara (2021), " Deep Learning Framework for Real-Time Estimation of in-silico Thrombotic Risk Indices in the Left Atrial Appendage ", Frontiers in Physiology, 28 de juny, doi.org/10.3389/fphy­s.2021.694945 .


This site uses cookies and analysis tools to improve the usability of the site. More information. |