PO-620-07 AN ENSEMBLE OF FEATURES BASED DEEP LEARNING NEURAL NETWORK FOR REDUCTION OF INAPPROPRIATE ATRIAL FIBRILLATION DETECTION IN IMPLANTABLE CARDIAC MONITORS

Abstract

Multiple studies have been reported for electrocardiogram (ECG) classification using deep learning neural networks. Most of these studies use the ECG as input to a convolution neural network (CNN) which then automatically generate features from the raw ECG signal. To develop an application specific deep learning CNNs using custom ensemble of features to reduce inappropriate AF detections in implantable cardiac monitors (ICM). The ensemble of features was developed and combined to form an input signal for deep learning CNN. The ensemble of features were designed based on the electro-physiological characteristics during AF: presence of fibrillation/flutter waves or absence of single p-wave between two R-waves, incoherence of RR intervals, and AF begetting more AF. A small custom CNN model using 6 convolution layers and the publicly available RESNET18 model were used. The deep learning models were trained and validated using more than 60K ICM detected AF episodes that were adjudicated to be true AF or false detections. The trained models were evaluated using an independent test dataset of ICM detected and adjudicated AF episodes from patients who were not included in the training and validation dataset. The training and validation dataset consisted of 31,768 true AF episodes (2516 patients) and 28,527 false episodes (2126 patients). The validation set used 20% of randomly chosen episodes of each type. The independent test set consisted of 4546 true AF episodes (418 patients) and 5384 false episodes (605 patients). In the validation set, the custom CNN had an area under ROC curve (AUC) of 0.996 (0.993 for RESNET18) and a threshold for discrimination was defined such that a relative sensitivity and specificity of 99.2% and 92.8% (99.2% and 87.9% for RESNET18) was obtained. The performance results for the two models in the independent test dataset for the pre-defined threshold is shown in the figure, with relative sensitivity and specificity of around 99% and 90% respectively. An ensemble of features based deep learning CNN was developed that reduced inappropriate AF detection in ICM by over 90% while preserving sensitivity for detection of true AF. For an application specific feature based CNN, a smaller sized custom CNN performed as well as a larger sized established CNN.