Sleep Disorder Analysis Using Large Language Models

Area of Science:

• Neuroscience
• Computational Biology
• Medical Informatics

Background:

• Electroencephalography (EEG) is crucial for sleep medicine, aiding in diagnosing sleep disorders.
• Manual sleep stage analysis is time-consuming, error-prone, and limited by data privacy.
• Existing machine learning models face scalability challenges with large EEG datasets.

Purpose of the Study:

• To investigate the efficacy of Large Language Models (LLMs) for classifying EEG-derived sleep stages.
• To evaluate BERT and Longformer models for distinguishing healthy versus unhealthy sleep patterns.
• To assess LLM performance on diverse EEG data segments.

Main Methods:

• Fine-tuning BERT and Longformer models on curated EEG datasets.
• Utilizing full-night sequences, two-hour segments, and mid-sleep windows for training.
• Classifying sleep stages as healthy or unhealthy based on EEG signals.

Main Results:

• LLMs achieved promising accuracies between 70-80% in classifying sleep stages.
• Demonstrated adaptability across different lengths of EEG data.
• Showcased potential for automated sleep disorder assessment.

Conclusions:

• LLMs offer a viable and scalable approach for automated sleep stage classification from EEG.
• This technology can enhance diagnostic capabilities in sleep medicine.
• Further research can explore LLMs for predicting sleep disorder progression.