Effectiveness of Transfer Learning for Deep Learning-Based Electrocardiogram Analysis
Problem
Individual researchers cannot collect the large labeled biosignal datasets needed to train deep learning models for electrocardiogram rhythm analysis.
Why it matters
Data scarcity blocks development of accurate automated ECG diagnostic tools for heart rhythm conditions.
What the AI did
A convolutional autoencoder was pretrained on 2,648,100 unlabeled ECG samples, then its weights were transferred to a classifier for 12 ECG rhythms using 10,646 labeled 12-lead ECGs, with performance compared against baselines.
Results
not stated in source
Impact
The study evaluates transfer learning as a way to overcome training-data scarcity, but the measured performance is not stated in the provided abstract.
Record
| Venue | Healthcare Informatics Research |
|---|---|
| Published | 2021 |
| Work type | article |
| Authors | Jong-Hwan Jang, Tae Young Kim, Dukyong Yoon |
| Citations | 43 (OpenAlex, at import time) |
| References | 24 |
| Model or method | convolutional autoencoder (CAE) with transfer learning |
| Organization | not stated in source |
| Primary source | https://doi.org/10.4258/hir.2021.27.1.19 |
| Open access | gold |
| Keywords | transfer learning, ECG, deep learning, convolutional autoencoder, rhythm classification |
Importance rating
real but narrower advance
Data scarcity for ECG deep learning is a real methodological barrier to automated cardiac diagnosis; abstract states researchers lack adequate training data but gives no scale of affected population.