Deep Learning for Amplified P-Wave Duration Annotation

Atrial cardiomyopathy (AtCM) is associated with new-onset atrial fibrillation (AF), higher AF recurrence rates after pulmonary vein isolation (PVI), and increased risk for ischemic stroke. Automated diagnosis of AtCM using electrocardiograms (ECGs) could enable non-invasive screening of large cohorts. The amplified P-wave duration (APWD) holds potential for diagnosing and staging AtCM. In this study, we propose a long short-term memory (LSTM) model to annotate APWD. The model’s training involved two phases: initial pretraining with weak labels and subsequent fine-tuning with expert labels. We investigated the effects of pretraining, trimming input signals, and upsampling on the absolute error between predictions and labels. The best-performing model was a bidirectional LSTM with 16 hidden units using pretraining, no trimming, and upsampling during fine-tuning, resulting in absolute errors of 13.9 ± 24.9, 15.4 ± 17.4, and 18.2 ± 19.8 ms for the P-wave onset, offset and duration, respectively. On the independent data set, errors were 7.3 ± 7.4, 15.6 ± 16.5, and 16.5 ± 21.1 ms, accordingly. The model showed little systematic bias and generalized well to unseen data. In conclusion, this work demonstrates promising results for the automation of AtCM diagnosis, suggesting potential for improved screening efficiency, ultimately enabling improved patient management and outcome.