Related Experiment Video
Updated: Oct 26, 2025
08:22
Author Spotlight: Advancing the Study of Brain-Heart Interplay with a Comprehensive EEGLAB Plugin for Multimodal Signal Analysis
Published on: April 26, 2024
2.4K
Cross-Gram matrices and their use in transfer learning: Application to automatic REM detection using heart rate
Bruno Muller1, Régis Lengellé2
1Institut Charles Delaunay, UTT, 12 rue Marie Curie, CS 42060, 10004 Troyes CEDEX, France; PPRS, 4E avenue du Général de Gaulle, 68000 Colmar, France.
Computer Methods and Programs in Biomedicine
|August 1, 2021
Summary
This study introduces novel transfer learning methods for automated sleep stage detection using heart rate, improving accuracy and efficiency in identifying Rapid-Eye-Movement sleep. These kernel-based approaches offer a practical alternative to traditional methods.
Area of Science:
- Machine Learning
- Computational Neuroscience
- Biomedical Signal Processing
Background:
- Traditional polysomnography for sleep staging is impractical and expensive.
- Automated methods using machine learning and heart rate (HR) show promise but face challenges due to physiological variabilities.
- Transfer learning is crucial for addressing data and concept shifts in HR-based sleep analysis.
Purpose of the Study:
- To propose and evaluate two novel kernel-based transfer learning methods for automated sleep stage detection.
- To assess the performance of these methods in identifying Rapid-Eye-Movement (REM) sleep using only heart rate.
- To introduce Kernel-Cross Alignment (KCA) as a new similarity measure between source and target domains.
Main Methods:
- Developed two alignment-based transfer learning methods: Kernel-Target Alignment Transfer Learning (KTATL) and Kernel-Cross Alignment Transfer Learning (KCATL).
- Introduced Kernel-Cross Alignment (KCA), an extension of Kernel-Target Alignment (KTA), as a measure of source-target similarity.
- Fixed the kernel beforehand and focused on improving target label estimation by considering inter-observation relationships for knowledge transfer.
Main Results:
- Both KTATL and KCATL demonstrated improved performance compared to a non-transfer learning approach (fixed Support Vector Classifier).
- The transfer learning methods achieved higher detection rates for a given false-alarm rate in REM sleep detection.
- The proposed methods are computationally efficient and do not require iterative calculations.
Conclusions:
- The novel kernel-based transfer learning methods (KTATL, KCATL) enhance the accuracy of automated REM sleep detection using heart rate.
- These methods are computationally efficient, relying on kernel matrix computation and avoiding iterative processes.
- Further optimization aspects of these transfer learning techniques are currently under investigation.
