Select for better learning: identifying high-quality training data for a multimodal cyclic transformer.

Area of Science:

• Epilepsy research
• Biomedical signal processing
• Machine learning in healthcare

Background:

• Tonic-clonic seizures (TCSs) pose a risk for sudden unexpected death in epilepsy (SUDEP).
• Accurate and reliable long-term monitoring of TCSs is crucial for patient management.
• Multimodal seizure detection systems show promise but depend on high-quality training data.

Purpose of the Study:

• To develop an innovative data selection method for identifying high-quality training samples for seizure detection models.
• To enhance the training pipeline for multimodal seizure detection systems.

Main Methods:

• Proposed a novel data selection approach evaluating sample quality based on learning difficulty.
• Classified samples with lower learning difficulty as higher quality.
• Introduced a confidence-based method to quantify high-quality samples within a dataset.

Main Results:

• The proposed data selection method improved the performance of a state-of-the-art TCS detection model by 11%.
• Demonstrated enhanced training process for multimodal seizure detection models.

Conclusions:

• The developed data selection method effectively enhances the training of multimodal seizure detection models.
• This approach contributes to more reliable and accurate long-term monitoring of tonic-clonic seizures.
• Improved data selection is key to advancing AI-driven epilepsy management tools.