Neural network-based prediction of incident atrial fibrillation using the trans-thoracic echocardiography parameters: a prospective study

Background: Myocardial remodeling facilitates the development and maintenance of atrial fibrillation (AF). Trans-thoracic echocardiography (TTE) is an available and easily reproducible method for assessment of patients with cardiovascular disorders. The evaluation of TTE parameters with a neural network-based program (NN) could be used for early detection of AF in asymptomatic patients. Aim: To establish the potential of a previously developed NN to predict the incident AF in the patients with no past history of heart arrhythmias. Methods: The study was based on the electrocardiographic data from electronic case record forms of 256 patients (mean age 63.47 ± 16.21 years, 151 (58.98%) women) with no past history of AF, who were treated in the in-patient Department of Cardiology in 2022. The TTE parameters (ascending aorta and left atrium diameters, left ventricular end-diastolic size, pulmonary artery diameter, maximal thickness of the anterior and posterior left ventricular walls in the diastole, transverse diameter of the right atrium, aortal, mitral and tricuspid valve regurgitation grade, left ventricular ejection fraction) were processed by the NN “Prediction of atrial fibrillation based on the trans-thoracic echocardiography data” (Russian state registration certificate for the computer software #2023662423 from 07.06.2023). Depending on the NN output values, the patients were divided into three groups: with a high probability of incident AF (n = 48; NN output value ≤ 0.33), with an intermediate probability of incident AF (n = 185, NN output value 0.34 to 0.66), and with a low probability (n = 23, NN output values ≥ 0.67). The patients were followed up to March 1, 2024 based on their electronic medical records (endpoints: referrals due to AF, AF mentioned as a complication of the underlying disease or as a concomitant disorder, cardiovascular death). Results: The median time of the follow-up was 16 [14; 21] months. During the study period, AF developed in 8/48 (16%) patients with the NN values ≤ 0.33 and in 4/185 (2.16%) patients with the NN values of 0.34 to 0.66, whereas no AF cases were registered in those with the NN values ≥ 0.67 (0/23). The cardiovascular death rate in the study groups was 17/48 (31.25%), 13/185 (3.78%), and 0/23 (0%), respectively (р 0.05). The area under the ROC curve (AUC) for incident AF was 0.85, for death 0.84, and for the combination endpoint of incident AF + death 0.86. Conclusion: The NN used in the study meets the goal of the binary differentiation, i. e. it allows to differentiate the heart structure and function typical for AF from those not typical for AF. Heart chamber dilatation, left ventricular systolic function and valvular regurgitation grade are associated with an increased risk of AF and death. The analysis of TTE parameters by the NN can be used for identification of patients at risk of incident AF with the goal of their additional work-up and monitoring.