Deep Learning-based Discrimination of Pause Episodes in Insertable Cardiac Monitors
Keywords: Insertable cardiac monitors, ECG, CNN, Remote monitoring, Pause detection, AI, Machine learning, Data augmentation.
TL;DR: A customized CNN model developed and tested in this study substantially reduced false pause episodes, with minimal impact on true pause episodes.
Abstract: Remote monitoring of patients with insertable cardiac monitors (ICMs) has revolutionized follow-up procedures and enhanced the timely diagnosis of cardiac arrhythmias. Despite these advancements, challenges persist in managing and adjudicating the data generated, placing strain on clinic resources. In response, various studies have explored the application of Convolutional Neural Networks (CNNs) to classify raw electrocardiograms (ECGs).
The objective of this study was to create and assess a CNN tailored for the reduction of inappropriate pause detections in ICMs. A customized end-to-end CNN model comprising 5 convolutional layers for rhythm classification of ICM-detected pause episodes was developed. The training data consisted of ICM-detected pause episodes from 1,173 patients. After training the model, we evaluated its performance using a test dataset of ICM-detected pause episodes from 750 independent patients. All pause episodes utilized in training and testing were adjudicated manually as either true or false detection. The training dataset consisted of 4,308 pause episodes (2,722 true episodes from 960 patients and 1,586 false episodes from 251 patients). The validation dataset includes 1,095 detected Pause episodes from 256 patients (677 true pause from 203 patients and 418 false pause episodes from 58 patients) and had an area under the curve (AUC) of 0.994 for the proposed CNN. The optimal threshold was chosen to obtain 99.26\% sensitivity and 96.89\%. The test dataset consisted of 1,986 episodes (744 true episodes from 382 patients and 1,242 false episodes from 485 patients. The model demonstrated an AUC of 0.9942, 99.06\% sensitivity, and 95.17\% specificity in the test dataset.
The customized CNN model, 737 out of 744 episodes were correctly identified as pauses, resulting in a positive predictive value of 92.47\%. Consequently, there was a reduction of EGM burden by 59.87\%.
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 6533
Loading