Sleep Disorder Analysis Using Large Language Models

Electroencephalography (EEG) recordings play a vital role in sleep medicine, offering critical insights into brain activity for diagnosing and predicting disorders like insomnia and sleep apnea. Traditionally, sleep disorder analysis relied on manual evaluation by clinicians using classified sleep-stage data, which is time-consuming, labor-intensive, and prone to errors. Additionally, privacy concerns restrict access to larger datasets, limiting publicly available data. While earlier machine learning methods showed accuracy on small datasets, they struggled with scalability for larger ones.This paper explores the use of Large Language Models (LLMs), specifically BERT and Longformer, to classify EEG-derived sleep stages as healthy or unhealthy. Fine-tuning is performed on curated datasets containing full-night sequences, two-hour segments, and mid-sleep windows. Results demonstrate a promising performance in classifying sleep stages with accuracies between 70-80%, showcasing their adaptability for sleep medicine.