Utilizing Pretrained Vision Transformers and Large Language Models for Epileptic Seizure Prediction

Area of Science:

• Artificial Intelligence
• Neuroscience
• Machine Learning

Background:

• Epilepsy seizure prediction is crucial for patient well-being.
• Challenges in seizure prediction include data variability and annotation complexity.
• Existing models require specialized knowledge and extensive data.

Purpose of the Study:

• To explore the efficacy of pre-trained AI models for seizure prediction.
• To compare the performance of Large Language Models (LLMs) and Vision Transformers (ViTs) in seizure prediction.
• To address the limitations of traditional supervised learning models in epilepsy research.

Main Methods:

• Utilized pre-trained Vision Transformers (ViTs) and Large Language Models (LLMs).
• Minimalistic refinement of input, embedding, and classification layers.
• Employed patient-independent seizure prediction methodology.

Main Results:

• LLMs outperformed ViTs in patient-independent seizure prediction.
• LLMs achieved a sensitivity of 79.02%, 8% higher than ViTs.
• LLMs showed a 12% improvement over a custom ResNet-based model.

Conclusions:

• Pre-trained LLMs are feasible and effective for seizure prediction.
• This approach offers a potential improvement in the quality of life for epilepsy patients.
• Open-source code is available for further research and development.