Researchers at the Polytechnic University of Valencia have developed an artificial intelligence model that can determine the location and extent of damage to atrial tissue using electrical signals recorded from the surface of the body. In an experiment, the system identified the location of affected areas with 89% accuracy and estimated the extent of the damage with 84% accuracy.
The work by researchers from the COR group at the Institute for Information and Communication Technologies (ITACA) was published in the journal Discover Computing. For the analysis, they used a graph neural network, a type of AI capable of accounting for relationships between different data points.
The study focused on atrial cardiomyopathy, a condition in which electrical and structural changes, including fibrosis, occur in the atrial tissue. Such changes are associated with the development and progression of atrial fibrillation, one of the most common heart rhythm disorders.
To assess the condition of atrial tissue, doctors can currently use invasive intracardiac electrocardiographic mapping and certain types of magnetic resonance imaging. The new method involves using body-surface potential maps: electrodes placed on the torso record the heart's electrical activity, after which the AI analyzes the spatial distribution of the signals and how they change over time.
To train and test the model, the researchers used 14,400 simulated recordings generated from different anatomical models of the atria and torso. With 128 electrodes, the system correctly identified the location of damaged tissue in 89% of cases. Sensitivity was 89%, while specificity was 90%.
The model also demonstrated an ability to assess the extent of tissue damage, achieving an overall accuracy of 84% in determining its spread. At the same time, the system maintained relatively stable performance even when the quality of the electrical signal was reduced, indicating a degree of robustness to noise.
The researchers consider it particularly important that the model remained functional when analyzing anatomical variations that had not been used during training. In addition, increasing the diversity of atrial and torso models in the training dataset improved classification performance.
However, the work is still only a proof of concept. All the data used in the study were simulated, so the results cannot yet be considered evidence that the method is effective in patients. The next step will be to test the system on real clinical recordings.
Follow NEWS.am Medicine on Facebook and Twitter