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 us- ing electrocardiograms (ECGs) could enable non-invasive screening of large cohorts. The amplified P-wave dura- tion (APWD) holds potential for diagnosing and staging AtCM. In this study, we propose a long short-term mem- ory (LSTM) model to annotate APWD. The model's train- ing involved two phases: initial pretraining with weak la- bels and subsequent training with expert labels. We inves- tigated 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 training, resulting in absolute er- rors 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 lit- tle systematic bias and generalized well to unseen data. In conclusion, this work demonstrates promising results for the automation of AtCM diagnosis, suggesting poten- tial for improved screening efficiency, ultimately enabling improved patient management and outcome.