Semi-Supervised Analysis of the Electrocardiogram Using Deep Generative Models.

Area of Science:

• Medical informatics
• Artificial intelligence in healthcare
• Cardiology

Background:

• Deep learning models achieve high performance in medical classification.
• Neural networks require substantial labeled data, which is costly and difficult to acquire for medical datasets.
• Existing methods often necessitate large labeled datasets for effective training.

Purpose of the Study:

• To develop and evaluate a semi-supervised learning approach for classifying atrial fibrillation using electrocardiogram (ECG) data.
• To compare the performance of the proposed semi-supervised model against a fully-supervised model.
• To assess the model's efficacy with varying proportions of labeled and unlabeled data.

Main Methods:

• A semi-supervised model combining an unsupervised variational autoencoder with a supervised classifier was proposed.
• The model was trained to distinguish between atrial fibrillation and non-atrial fibrillation using the MIT-BIH Atrial Fibrillation Database.
• Performance was evaluated against a fully-supervised convolutional neural network using 1%-50% labeled data.

Main Results:

• The semi-supervised approach demonstrated superior performance compared to the fully-supervised model.
• High accuracy (98.7%) was achieved using only 5% labeled data (5,594 samples).
• The model maintained high accuracy even with significantly reduced labeled data proportions.

Conclusions:

• The proposed semi-supervised learning setup effectively trains high-accuracy atrial fibrillation detection models with minimal labeled ECG data.
• This approach offers a viable solution for overcoming data scarcity in medical deep learning.
• The study provides proof of concept for efficient data utilization in medical AI.