top of page
Search

Using AI to Revolutionize A-Fib Treatment: Insights from Focus Theory and "Tailored AF" Trials

This blog post serves to simplify these scholarly papers:


What is Atrial Fibrillation (AF)?

Atrial Fibrillation is one of the most common heart diseases, only trumped by heart attacks and heart failures.


The heart beats are orchestrated by electrical signals sent out from the atria. These signals help the muscle fibers fire at the same time allowing for a uniform pump. 


In the case of Atrial Fibrillation, regions in the atria called “dispersion regions” fire chaotic electrical signals causing the muscle fibers to fire out of sync, a phenomena called ectopic heartbeats. This is what causes the characteristic “Fibrillation” or quivering of the heart. (See Figure 1 to help make sense of these words).


How do dispersion regions come about? 

The electrical signals of the heart start in the SA node, the master coordinator of the heart beat. The electrical signal propagates through the heart muscles and allows for a uniform pump. In AFib however, some muscle tissue is damaged due to cardiac remodeling. This causes the electrical signal to propagate abnormally through that muscle tissue, in turn causing downstream muscle fibers to fire abnormally too.


Cardiac remodeling occurs when the cells in a certain area deteriorate. Over time, risk factors like high blood pressure, obesity, heart disease, or sleep apnea put the atria under stress. This stress can:

  • Cause muscle cells to enlarge, weaken, or die.

  • Lead to scar tissue (fibrosis) that blocks or slows electrical conduction.

  • Alter the extracellular matrix, changing how cells connect and communicate.


These structural changes create uneven conduction paths. Even if the SA node sends a normal signal, the damaged tissue can distort it — some areas conduct quickly, others slowly, and some not at all. This electrical mismatch is what creates dispersion regions, allowing chaotic reentrant circuits to form and sustain AFib.


Flowchart titled "Simplified Understand of the A-Fib Pathology" shows pathways of electrical signals in heart tissue, leading to normal or abnormal heartbeat.
Figure 1: This diagram shows how risk factors eventually lead to A-Fib. It also shows how cardiac remodeling and the formation of dispersion regions fit into the larger picture.

Living with AF could lead to severe fatigue, heart failure, and even death, so what are the treatment options?


Traditional AF Treatment Plans

Today’s gold standard for AF treatment includes a catheter ablation (ablation means to destroy) of the pulmonary vein, a procedure that creates scars in the heart muscle tissue to block off or isolate chaotic electrical signals coming from the pulmonary vein. 

The reason for targeting the pulmonary vein is that it is the most common dispersion region. In simpler terms, the tissues around the pulmonary vein are well-known to fire chaotic electrical signals in high risk patients, and Pulmonary Vein Isolation (PVI) procedure tries to stop this.


And it works, usually. 

Although the pulmonary veins are a well-established source of abnormal electrical firing, patients with long-standing AF often have various other dispersion regions that, left untreated, will land them in the hospital again and again. 


Mapping out each patient's unique electrical landscape using Machine Learning models is the next step.


AI for A-Fib:

The researchers behind the article published in Nature Medicine: "Artificial intelligence for individualized treatment of persistent atrial fibrillation: a randomized controlled trial" compared two groups: The first group received a traditional PVI procedure for their long-standing A-Fib. The second group would take part in Tailored A-Fib ablation procedure where patient data was ran through an AI model that detected dispersion areas other than the pulmonary vein.

"The purpose of the TAILORED-AF randomized controlled trial was to evaluate whether a tailored cardiac-ablation procedure targeting AI-detected areas harboring spatio-temporal dispersion, in addition to PVI, is more effective than an anatomical PVI-only procedure in patients with persistent and long-standing persistent AF."

Their research question boils down to: How do patients who undergo PVI + AI detected area isolation procedures compare to those patients who stick to the traditional PVI only practice?


What was their success criteria?

Any patient, whether from the traditional group or the AI-detected dispersion area group, would be considered a successful patient if they were free from AF at the 12 month mark after a single ablation procedure.


Did it work?

Yes, 88% of the AI-guided treatment group was free from AF after one year. Compare that to the 70% of the Standard treatment group who were AF free after one year.


Making sense of the Machine-Learning algorithm used:

To make sense of any ML model identify four things:

  • What type of model this is? (binary classification model)

  • What was the input data? (EGM data for each heart region)

  • What algorithm was used on the input data? (two algorithms types were used, XGBoost, and a convolutional neural network)

  • What was the output data? (Colored 3D map of dispersion areas)


Volta AF-Xplorer, the ML model behind this massive advance. The fundamental thing to understand about this model to understand is that it's job is to classify. More specifically, it does binary classification, which means it classifies input data into one of two categories: normal or abnormal. That answers the first of the four questions.

The input data for Volta AF-Xplorer is the EGM wave time-series data for each electrode site. This means that the electrical activity for each area of the heart is recorded in the form of a wave that is fed to the ML model.

Before we talk about which algorithms were used, what even are algorithms? Algorithms are the rules that the model will follow to come to an output. One model may follow handcrafted rules that a human designed, while another model has more freedom to learn and come up with its own rules. Volta AF-Xplorer does both. It uses two models, one called XGBoost that is better with numbers, and another Convolutional Neural Network, which works with the shape and soft features of the waveform.

Lastly, what are the outputs? Each algorithm outputs a "probability score" out of one. These scores are combined in a "Fusion" step into a "final probability score", also out of one. If this score is, for example, 0.7, and the threshold is 0.5, we can say that this specific electrode was connected to a dispersion positive site. This data of dispersion-positive or normal is fed into a mapping system which ultimately yields the 3D colored dispersion area map of the atria.

Figure 2: The full EGM to 3D map workflow
Figure 2: The full EGM to 3D map workflow


More Areas of Research:

  • AI-guided ablation in paroxysmal AF

  • Pulsed field ablation (PFA) + AI mapping

  • Machine learning for AFib recurrence prediction

  • Electrogram-based dispersion mapping

  • AI vs. operator variability in AF ablation

  • Integrating AI with EnSite X and CARTO mapping systems



More Resources:

 
 
bottom of page