Publication detail

ECG Abnormalities Recognition Using Convolutional Networkwith Global Skip Connections and Custom Loss Function

VIČAR, T. HEJČ, J. NOVOTNÁ, P. RONZHINA, M. JANOUŠEK, O.

Original Title

ECG Abnormalities Recognition Using Convolutional Networkwith Global Skip Connections and Custom Loss Function

English Title

ECG Abnormalities Recognition Using Convolutional Networkwith Global Skip Connections and Custom Loss Function

Type

conference paper

Language

en

Original Abstract

The latest trends in clinical care and telemedicine sug-gest a demand for a reliable automated electrocardiogram(ECG) signal classification methods. In this paper, wepresent customized deep learning model for ECG classi-fication as a part of the Physionet/CinC Challenge 2020.The method is based on modified ResNet type convolu-tional neural network and is capable to automatically rec-ognize 24 cardiac abnormalities using 12-lead ECG. Wehave adopted several preprocessing and learning tech-niques including custom tailored loss function, dedicatedclassification layer and Bayesian threshold optimizationwhich have major positive impact on the model perfor-mance. At the official phase of the Challenge, our team -BUTTeam - reached a challenge validation score of 0.696,and the full test score of 0.202, placing us 21 out of 40in the official ranking. This implies that our method per-formed well on data from the same source (reached firstplace with validation score), however, it has very poor gen-eralization to data from different sources.

English abstract

The latest trends in clinical care and telemedicine sug-gest a demand for a reliable automated electrocardiogram(ECG) signal classification methods. In this paper, wepresent customized deep learning model for ECG classi-fication as a part of the Physionet/CinC Challenge 2020.The method is based on modified ResNet type convolu-tional neural network and is capable to automatically rec-ognize 24 cardiac abnormalities using 12-lead ECG. Wehave adopted several preprocessing and learning tech-niques including custom tailored loss function, dedicatedclassification layer and Bayesian threshold optimizationwhich have major positive impact on the model perfor-mance. At the official phase of the Challenge, our team -BUTTeam - reached a challenge validation score of 0.696,and the full test score of 0.202, placing us 21 out of 40in the official ranking. This implies that our method per-formed well on data from the same source (reached firstplace with validation score), however, it has very poor gen-eralization to data from different sources.

Keywords

ECG, arrhythmia, signal, classification, challenge, CinC 2020, Computing in Cardiology

Released

30.09.2020

Publisher

Computing in Cardiology 2020

Location

Rimini, Italy

ISBN

0276-6574

Periodical

Computers in Cardiology

State

US

Pages from

1

Pages to

4

Pages count

4

Documents

BibTex


@inproceedings{BUT165485,
  author="Tomáš {Vičar} and Jakub {Hejč} and Petra {Novotná} and Marina {Ronzhina} and Oto {Janoušek}",
  title="ECG Abnormalities Recognition Using Convolutional Networkwith Global Skip Connections and Custom Loss Function",
  annote="The latest trends in clinical care and telemedicine sug-gest a demand for a reliable automated electrocardiogram(ECG)  signal  classification  methods.   In  this  paper,  wepresent customized deep learning model for ECG classi-fication as a part of the Physionet/CinC Challenge 2020.The  method  is  based  on  modified  ResNet  type  convolu-tional neural network and is capable to automatically rec-ognize 24 cardiac abnormalities using 12-lead ECG. Wehave  adopted  several  preprocessing  and  learning  tech-niques including custom tailored loss function, dedicatedclassification  layer  and  Bayesian  threshold  optimizationwhich  have  major  positive  impact  on  the  model  perfor-mance.  At the official phase of the Challenge, our team -BUTTeam - reached a challenge validation score of 0.696,and the full test score of 0.202,  placing us 21 out of 40in the official ranking.  This implies that our method per-formed well on data from the same source (reached firstplace with validation score), however, it has very poor gen-eralization to data from different sources.",
  address="Computing in Cardiology 2020",
  booktitle="Computing in Cardiology 2020",
  chapter="165485",
  edition="47",
  howpublished="online",
  institution="Computing in Cardiology 2020",
  year="2020",
  month="september",
  pages="1--4",
  publisher="Computing in Cardiology 2020",
  type="conference paper"
}