How Much Data is Enough? Benchmarking Transfer Learning for Few Shot ECG Image Classification

Sathvik Bhaskarpandit

2022

Abstract

Over the past couple of decades, numerous research works have been conducted to study and detect abnormalities from ECG signals. In this direction, several deep learning models have been proposed to detect these abnormalities and aid healthcare experts in their diagnoses. Although many of these deep learning approaches utilize ECG signals as input, only a handful use images of patients’ ECGs themselves, that are often stored in hospitals and diagnostic centres. This work aims to study ECG images under the few-shot learning scenario. More specifically, it aims to study the effectiveness of transfer learning for few-shot ECG image classification, and how classification performance varies with the amount of training data available. Results show that models such as ResNet and EfficientNet are able to classify images with great success with around 20 images per class, with accuracy even crossing 99.5%. Yet under extreme data unavailability cases, such as 5-shot learning and lower, transfer learning proves to be unreliable to be put to use in healthcare for automated classification of ECG images.

Paper Citation

in Harvard Style

Bhaskarpandit S. (2022). How Much Data is Enough? Benchmarking Transfer Learning for Few Shot ECG Image Classification. In Proceedings of the 1st Workshop on Scarce Data in Artificial Intelligence for Healthcare - Volume 1: SDAIH, ISBN 978-989-758-629-3, SciTePress, pages 35-40. DOI: 10.5220/0011531700003523

in Bibtex Style

@conference{sdaih22,
author={Sathvik Bhaskarpandit},
title={How Much Data is Enough? Benchmarking Transfer Learning for Few Shot ECG Image Classification},
booktitle={Proceedings of the 1st Workshop on Scarce Data in Artificial Intelligence for Healthcare - Volume 1: SDAIH,},
year={2022},
pages={35-40},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0011531700003523},
isbn={978-989-758-629-3},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 1st Workshop on Scarce Data in Artificial Intelligence for Healthcare - Volume 1: SDAIH,
TI - How Much Data is Enough? Benchmarking Transfer Learning for Few Shot ECG Image Classification
SN - 978-989-758-629-3
AU - Bhaskarpandit S.
PY - 2022
SP - 35
EP - 40
DO - 10.5220/0011531700003523
PB - SciTePress