Atrial Fibrillation Detection with CNNs for Insertable Cardiac Monitors

Multiple studies have reported on classification of raw electrocardiograms (ECGs) using convolution neural networks (CNN). Most of these models try to classify for all kinds of arrhythmia from conventional ECGs. Discover how Medtronic investigated an application-specific CNN using a custom ensemble of features designed with characteristics of the ECG during atrial fibrillation (AF) to reduce inappropriate AF detections in implantable cardiac monitors (ICM). The ensemble of features was developed and combined to form a 2D input signal for a CNN. They were based on the morphological characteristics of AF, incoherence of RR intervals, and the fact that AF begets more AF. A custom CNN model (similar to ones used in MATLAB^® examples) and the RESNET-18 model were trained using ICM detected AF episodes. The training, validation, and independent test data sets were created using real-world ICM detected episode data in the Medtronic data warehouse that were manually adjudicated to be true AF or false detections. The trained models were evaluated using a test data set from independent patients using convention sensitivity, specificity, and ROC curves. The training and validation data set consisted of 31,757 AF episodes (2,516 patients) and 28,506 false episodes (2,126 patients). The validation set (20% randomly chosen episodes of each type) had an AUC of 0.996 for custom CNN (0.993 for RESNET-18). Thresholds were chosen to obtain a relative sensitivity and specificity of 99.2% and 92.8% respectively (99.2% and 87.9% for RESNET-18). The performance in the independent test set (4,546 AF episodes from 418 patients; 5,384 false episodes from 605 patients) showed an AUC of 0.993 (0.991 for RESNET-18) and relative sensitivity and specificity of 98.7% and 91.4% respectively at chosen thresholds (98.9% and 88.2% for RESNET-18). In conclusion, the ensemble of features–based CNN reduced inappropriate AF detection in ICM by over 90% while preserving sensitivity.