Atrial fibrillation (AF) is one of the most common heart pathologies in daily clinical practice and confers significant mortality and morbidity. However, AF ablation often proves ineffective - even after repeated treatments. The prediction of ablation failure for the long-term treatment of AF as well as the improved assessment of the underlying substrate appears pivotal. In a recent study Attia et al. successfully applied Machine Learning (ML) to predict the development of AF from ordinary 12-lead electrocardiograms (The Lancet, 2019). We hypothesize, that ML can be applied to better characterize the substrate and recurrence in AF patients.

We retrospectively evaluated data from 134 patients that underwent AF ablation including high-density 3D electroanatomic maps (CARTO, sinus rhythm, low voltage was defined <0.5 mV), LA rotational angiography, invasive LA hemodynamic measurements (after transseptal passage), echocardiography, 12-lead ECG and laboratory testing at the German Heart Center in Berlin between 2019 and 2021.

Patients were 65±1 years old and received first or second pulmonary vein isolation (PVI) for paroxysmal atrial fibrillation with 19% documented recurrences upon a mean follow-up of 25±1months. First we extensively refined or developed tools for automated P-wave analysis from 12-lead ECGs as well as local LA voltage quantification. Next, LA hemodynamics and volumes were assessed: Mean LA volume as derived from rotational angiography significantly correlated with LA volume index as obtained using echocardiography. Interestingly, LA size showed no correlation with NT-proBNP, LA low voltage or LA hemodynamics. In addition global (21±5% of total LA area) and local (i.e. anterior, posterior, roof) LA voltage was determined. LA voltage and LA pressure significantly correlated with NT-proBNP upon a certain threshold, while LA pressure was independent of the extend of global LA low voltage. LA pressure A waves correlated with NT-proBNP. ML was employed to predict the left atrial volume and current adaptions of our supervised learning algorithm are trained to predict LA pressure and fibrosis in our set of patients.

Our project provides proof of principle evidence for ML as a tool for quantitative and qualitative atrial cardiomyopathy characterization (i.e. LA volume, pressure, fibrosis). Further training of artificial neural networks using our set of pre-processed data and supervised learning might allow to quantitatively and qualitatively predict additional measures of atrial cardiomyopathy and atrial fibrillation recurrence probability.