Deep learning-enabled ECG fingerprinting of low-voltage substrate in atrial fibrillation patients

Atrial fibrillation (AF) is a major and growing global health burden. Yet the optimal ablation targets that determine procedural success are still debated, with no consensus on how best to delineate the underlying atrial substrate. Low-voltage areas (LVAs), a common surrogate of atrial fibrosis, strongly predict sub-optimal outcomes after catheter ablation, but their identification remains confined to invasive electro-anatomical mapping. Prior ECG-based efforts focused on single P-wave metrics and yielded inconsistent, operator-dependent results. We hypothesized that domain-tailored DL strategies leveraging transfer learning from extensive public data to small datasets can uncover latent substrate signatures of ECG to predict LVAs.

To develop and validate a transfer deep-learning model that non-invasively predicts LVAs with a high degree of granularity in patients with paroxysmal and persistent AF from digital preprocedural 10 sec 12 lead ECG in sinus rhythm.

Sinus-rhythm ECGs and intra-procedural high-resolution electro-anatomical maps using a octaray catheter from N=132 consecutive patients (37.6% female, age=67±11 years, BMI=26.8±4.5Kg/m2) were retrospectively analyzed. For each AF patient (60.6% paroxysmal), the left atrium was manually segmented and anatomically labeled by domain experts to define regional ground-truth LVAs. LVAs were quantified using a bipolar voltage of <0.5 mV and a minimum area of ≥5 cm² across the following regions: posterior wall (PW), atrial roof (AR), anterior wall (AW) and myocardial isthmus (MI). We designed a densely connected convolutional network based on state-of-the-art ECG classification literature, and integrated recently proposed architectures. The model was first pre-trained on >2 million ECGs from public datasets using a self-supervised strategy, and then fine-tuned on our private dataset. Five-fold cross-validation on a development cohort optimized operating thresholds; final performance was assessed on an unseen held-out test set. AUROC, accuracy and precision were calculated against the binary reference of any LVA meeting the defined criteria in each anatomical region.

On the held-out test set, the model achieved an AUROC of 0.81 and accuracy of 0.73 for detecting the presence of any LVA. In the more fine-grained analysis, the model predicted PW, AR, AW, and MI the AUROCs were 0.76, 0.53, 0.65, and 0.71 respectively, with accuracies of 0.68, 0.80, 0.67, and 0.72.

A transfer-learning DenseNet applied to routine 12-lead ECGs identified LVAs with good discrimination across left-atrial segments. Used as an upfront screen, the model could help operators anticipate fibrotic substrate, inform mapping priorities, and potentially shorten ablation time. Prospective studies that link ECG-predicted LVAs to procedural modifications and long-term rhythm outcomes are needed before the approach can be integrated into routine clinical decision-making.

Study Pipeline and Method