Weakly Supervised P Wave Segmentation in Pathological Electrocardiogram Signals Using Deep Multiple-Instance Learning.

Detection of obscured P waves remains a largely unexplored topic. This study proposes a weakly supervised learning approach for P wave feature embedding by leveraging surrogate labels and 3265 eight-lead electrocardiographic (ECG) signals with diverse cardiac rhythms, including supraventricular tachycardias, atrial fibrillation, and paced rhythms. The proposed method employs a temporal convolutional neural network and multiple instance learning to learn pyramidal feature embeddings that estimate both labeled and unlabeled instances of the P wave. The fine-tuned model achieved a temporally aggregated Dice score of 81.1%, outperforming the baseline model by 1.0%. On the subset with sinus rhythms and minor rhythm irregularities, the model consistently achieved recall and precision of around 84-85% for P wave onset and offset. The framework can be used to learn embeddings correlated with the distribution of the atrial depolarization, using only a fraction of labeled samples. Surrogate labels allow us to embed more detailed context, which may enhance the performance and interpretability of deep neural networks in downstream tasks in the future.