The Resnet model achieved an f1-score of 99.0% for true AT/AF detection, significantly reducing false positive alarms in pacemaker and ICD remote monitoring.
Does a residual network deep learning model accurately classify AT/AF alerts from pacemakers and ICDs compared to expert review?
A deep learning residual network can accurately classify true AT/AF episodes from noise and oversensing in remote monitoring of pacemakers and ICDs, potentially reducing clinician workload.
Absolute Event Rate: 0% vs 0%
Abstract Background Remote monitoring (RM) of patients with pacemakers and implantable cardiac defibrillators (ICDs) allows quicker detection of device and health related complications compared to conventional follow-up. RM reduces hospitalization and mortality of patients, but increases the work burden on the health care professional. False positive (FP) arrhythmia alarms of pacemakers and ICDs further complicate RM. Deep learning models can help healthcare professionals by automatically reviewing RM episodes, filter out FP alarms, and improve patient security. Purpose To create a deep learning model which automatically classifies Atrial Tachycardia or Atrial Fibrillation (AT/AF) alerts of Biotronik pacemakers and ICDs into true AT/AF, noise, and oversensing alarms. Methods A residual network (Resnet) is designed to classify AT/AF episodes using atrial and ventricular intracardiac electrograms of AT/AF episodes, provided by three hospitals (80% train set, 20% validation set). The optimal combination of different layers, structures, and hyperparameters of the Resnet is determined by grid search and manual tuning. The resulting Resnet is trained with a focus on maintaining high true AT/AF sensitivity and is externally tested on a test set supplied by a fourth hospital. Lastly, real-time clinical validation was conducted using a chrome plug-in for two weeks in a hospital, as shown in figure 1. The performance of the model is quantified using the f1-score, to provide a balance between the positive predictive value and sensitivity, and adequately assess model performance on imbalanced data. All episodes are blind and independently labelled by two RM experts, a third RM expert settled conflicts between the two RM experts. Results The Resnet consists of a residual block (of 128 filters of length 640) and max pooling layer, four residual blocks for feature learning (with 128, 128, 64, 64 filters of length 640, 640, 320, 320, respectively), a global pooling layer, and two dense layers. The train/validation set consisted of 8892 episodes of 911 patients, the test set consisted of 1858 episodes of 237 patients, and the clinical data of 307 episodes of 68 patients. The train/validation and test set had similar distributions of true AT/AF (86%), noise (13%), and oversensing (1%). The Resnet model identified true AF episodes, noise, and oversensing alarms of the test set with a f1-score of 99.1%, 96.0%, and 73.3%, respectively. The clinical data consisted of 291 AT/AF (94.8%), 12 noise (3.9%), and 4 oversensing (1.3%) episodes. The Resnet achieved f1-scores of the 99.0%, 73.7%, and 88.9%, for the AT/AF, noise, and oversensing episodes, respectively. Conclusion The Resnet accurately identifies true AT/AF, noise, and oversensing as the root-cause behind the AF alerts in Biotronik pacemakers and ICDs. The ability of this model to detect true AT/AF with a high f1-score encourages future usage of this Resnet in hospitals to reduce the RM work burden.
Krimpen et al. (Thu,) reported a other. The Resnet model achieved an f1-score of 99.0% for true AT/AF detection, significantly reducing false positive alarms in pacemaker and ICD remote monitoring.