An ensemble of features based deep learning neural network for reduction of inappropriate atrial fibrillation detection in implantable cardiac monitors.

Background: Multiple studies have reported on classification of raw electrocardiograms (ECGs) using convolutional neural networks (CNNs). Objective: We investigated an application-specific CNN using a custom ensemble of features designed based on characteristics of the ECG during atrial fibrillation (AF) to reduce inappropriate AF detections in implantable cardiac monitors (ICMs). Methods: An ensemble of features was developed and combined to form an input signal for the CNN. The features were based on the morphological characteristics of AF, incoherence of RR intervals, and the fact that AF begets more AF. A custom CNN model and the RESNET18 model were trained using ICM-detected AF episodes that were adjudicated to be true AF or false detections. The trained models were evaluated using a test dataset from independent patients. Results: The training and validation datasets consisted of 31,757 AF episodes (2516 patients) and 28,506 false episodes (2126 patients). The validation set (20% randomly chosen episodes of each type) had an area under the curve of 0.996 for custom CNN (0.993 for RESNET18). Thresholds were chosen to obtain a relative sensitivity and specificity of 99.2% and 92.8%, respectively (99.2% and 87.9% for RESNET18, respectively). The performance in the independent test set (4546 AF episodes from 418 patients; 5384 false episodes from 605 patients) showed an area under the curve of 0.993 (0.991 for RESNET18) and relative sensitivity and specificity of 98.7% and 91.4%, respectively, at chosen thresholds (98.9% and 88.2% for RESNET18, respectively). Conclusion: An ensemble of features-based CNNs was developed that reduced inappropriate AF detection in ICMs by over 90% while preserving sensitivity.