1D Convolutional Neural Networks for Detecting Atrial Fibrillation

EasyChair Preprint 2831

Abstract

This work is a part of the mobile health monitoring system project in Sultan Qaboos' university, Muscat Oman. We explain in this work an effective and precise method of detecting Atrial Fibrillation from a single channel short electrocardiogram (ECG). The used ECG signals are downloaded from the Physionet/Computing in Cardiology Challenge 2017. Signals lengths varies between thirty and ninety seconds. The outputs are 3 different classes, Atrial Fibrillation (AF) Normal (N) and Noisy (∼). The proposed model is based on a deep learning one dimensional Convolutional Network, eliminating the need to manually extract features. R-peaks are detected using python's BioSPPy library then R to R intervals are calculated, stacked into a dataframe, amputated and parsed with a manually chosen value then injected into the neural network. The RR records are classified next into one of the three classes. The proposed model has reached 98% training accuracy, 96% validation accuracy and 94.07% testing accuracy.

Keyphrases: ECG, biomedicine, e-health, machine learning, signal processing

@booklet{EasyChair:2831,
  author    = {Khalil Ben Kalboussi and Lazhar Khriji and Farah Barika Ktata and Abdulnasir Hossen},
  title     = {1D Convolutional Neural Networks for Detecting Atrial Fibrillation},
  howpublished = {EasyChair Preprint 2831},
  year      = {EasyChair, 2020}}