KL Divergence-based transfer learning for cross-subject eye movement recognition with EOG signals

Area of Science:

• Biomedical Engineering
• Signal Processing
• Human-Computer Interaction

Background:

• Electrooculogram (EOG) signals are widely used for human-machine interfaces (HMIs).
• Cross-subject variability poses a significant challenge for EOG-based eye movement recognition models.
• Existing transfer learning methods often fail to account for subject-specific similarities.

Purpose of the Study:

• To develop an adaptive transfer learning framework for EOG-based eye movement recognition.
• To improve the performance of EOG models across different subjects.
• To address the limitations of traditional transfer learning methods in handling subject variability.

Main Methods:

• Utilized Kullback-Leibler (KL) divergence of log-Power Spectral Density (log-PSD) features of horizontal EOG (HEOG).
• Adaptively selected source subjects with distributions similar to the target subject.
• Implemented a novel transfer framework to enhance cross-subject model training.

Main Results:

• The proposed approach significantly outperformed baseline and classical transfer learning methods.
• Achieved substantial performance improvements for subjects with initially poor classifier results.
• Demonstrated a 13.1% improvement for a specific subject using a Support Vector Machines (SVM) classifier.

Conclusions:

• The developed transfer framework effectively enhances cross-subject EOG eye movement recognition.
• The method offers a promising solution for real-world applications requiring robust EOG-based HMIs.
• Adaptive subject selection based on feature distribution similarity is crucial for improving model generalization.