Finetuning enables Atrial Fibrillation Detection in 12-Lead Electrocardiograms with Active Pacing

Detecting atrial fibrillation (AF) in 12-lead electrocardiograms (ECGs) is challenging during active pacing as pacemaker impulses can mask typical AF features such as irregular heartbeats and atrial activity. This also leads to data sparsity with very limited training data being available. To address this, we adopted a state-of-the-art deep neural network which we finetuned using the recently published Harvard-Emory ECG Database. Four strictly matched subgroups were defined: AF with and without active pacing, and non-AF with and without active pacing. Hyperparameters were optimized through a systematic search involving five-fold cross-validation and data augmentation. The best model achieved an area under the curve of 0.899 on the validation set. External evaluation was performed using an independent subset of MIMIC-IV-ECG. In paced rhythms, our model improved performance substantially in terms of both accuracy (+22.8%) and F1 score (+15.9%) compared to the state-of-the-art model. Additionally, three cardiologists conducted a blinded review of 200 randomly sampled paced recordings. The majority vote achieved a sensitivity of 57.9% and specificity of 98.0%, whereas the model achieved 87.4% and 78.6%, respectively. Post-hoc explanations indicated physiologically meaningful feature shifts due to model finetuning, and model similarity analysis via centered kernel alignment showed that finetuning modified deeper layers. In conclusion, we demonstrate for the first time robust AI-based AF detecting in ECGs during active pacing.