Generalizable Seizure Prediction With LLMs: Converting EEG to Textual Representations.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Scalp electroencephalogram (EEG) based seizure prediction shows promise but faces challenges with patient heterogeneity and varying channel settings.
• Existing models struggle with generalizability due to individual differences and inconsistencies in data acquisition across epilepsy centers.
• Developing robust and adaptable seizure prediction models is crucial for clinical applications.

Purpose of the Study:

• To propose a novel seizure prediction method based on large language models (LLMs) to enhance model generalizability and applicability.
• To address the limitations of patient heterogeneity and diverse channel settings in current seizure prediction algorithms.
• To leverage the cross-domain knowledge and temporal dependency capture capabilities of LLMs for improved EEG analysis.

Main Methods:

• Reprogramming large language models (LLMs) by converting EEG signals into textual representations using a single-channel pre-training strategy.
• Integrating cross-domain knowledge from text and EEG data via a cross-attention mechanism.
• Utilizing autoregressive pre-trained LLMs to capture temporal dependencies in EEG signals for seizure prediction.

Main Results:

• The proposed seizure prediction method based on LLMs (SPLLM) significantly enhances model generalizability and applicability.
• SPLLM demonstrated an average increase of 8.2% in AUC and 8.4% in balanced accuracy across multiple datasets compared to existing methods.
• The method effectively alleviates patient heterogeneity and adapts to diverse channel settings, improving cross-patient prediction accuracy.

Conclusions:

• The SPLLM method offers a scalable solution for clinical seizure prediction by improving cross-patient accuracy and adaptability to different datasets.
• Incorporating LLMs into EEG analysis represents a significant advancement in seizure prediction technology.
• The proposed approach overcomes key limitations of existing methods, paving the way for more reliable seizure forecasting systems.