Unsupervised Feature Selection-Driven Active Learning for Semi-Supervised Automatic ECG Analysis.

Area of Science:

• Cardiology
• Machine Learning
• Biomedical Signal Processing

Background:

• Automatic electrocardiogram (ECG) analysis demands extensive annotated data, making manual annotation time-consuming.
• Semi-supervised learning (SSL) leverages unlabeled data but relies heavily on initial labeled subset quality.
• Existing active learning methods for ECG analysis face limitations in manual intervention, computational cost, and SSL compatibility.

Purpose of the Study:

• To develop an automated framework for ECG analysis that minimizes annotation requirements.
• To address the limitations of conventional active learning in the context of semi-supervised learning.
• To enhance the efficiency and accuracy of ECG analysis tasks through an integrated approach.

Main Methods:

• Proposed an Unsupervised Active Feature-selective Semi-Supervised Learning (UAFSSL) framework.
• Integrated unsupervised feature extraction for capturing latent data distributions.
• Employed pseudo-label clustering for selecting diverse and representative samples, eliminating human intervention.

Main Results:

• Improved F1-score by 2.4% for P-wave delineation in ECG segmentation compared to random sampling (using 5% labeled data).
• Achieved AUC improvements of 2.5% and 2.2% over random sampling for atrial fibrillation detection (using 200 labeled samples).
• Demonstrated robust performance on both AFDB and a 24-hour patient dataset.

Conclusions:

• UAFSSL offers a novel integration of unsupervised active learning and semi-supervised learning for automated ECG analysis.
• The framework significantly reduces annotation costs and enhances model performance.
• Presents a robust, automated solution with improved clinical applicability for ECG interpretation.