1D Convolutional Neural Network for Detecting Ventricular Heartbeats

Abstract

This paper shows a novel approach for detecting ventricular heartbeats using a 1D Convolutional Neural Network (1D-CNN). The algorithm input is the raw ECG signal, i.e., no signal pre-processing nor feature extraction are involved. The output of the 1D-CNN is filtered using a combination of linear and nonlinear filters to produce the final output. The MIT-BIH arrhythmia database was used for both algorithm training/tuning and evaluation. The assessment methodology followed the inter-patient paradigm, where the algorithm was trained and evaluated using independent subsets. The performance of the proposed method was evaluated for two tasks; QRS detection, and heart-beat classification. QRS detection resulted in a sensitivity of 99.0% and a positive predictivity of 96.5%. The performance assessment of the ventricular ectopic beat detection resulted in a sensitivity of 85.8% and a positive predictivity of 64.5%. Although there is still room for improvement, the results suggest that convolutional neural networks are a promising approach for building heartbeat classifiers.