Sleep Disorder Classification Using Bidirectional LSTM Models on Polysomnographic Signal Streams

This study tackles the essential need for precise, automated classification of sleep disorders using continuous polysomnographic data. The aim is to create a comprehensive deep learning framework that can differentiate among various sleep disorders by examining temporal signal patterns. Bidirectional Long Short-Term Memory (BiLSTM) models were used on the Sleep-EDF Expanded Dataset, including 197 complete polysomnographic recordings from entire nights, which include EEG, EOG, and EMG data, together with carefully annotated hypnograms. The proposed BiLSTM architecture effectively captures both forward and backward dependencies in sleep stages, hence improving the extraction of temporal context across successive sleep cycles. The model attained an accuracy of 97.57%, precision of 91.80%, recall of 91.16%, and F1-score of 91.48%, surpassing conventional machine learning methods used in similar tasks. The findings illustrate the efficacy of BiLSTM in acquiring profound temporal information from unprocessed sleep data streams. This research has practical applications in real-time clinical decision assistance, automated sleep monitoring, and home-based sleep diagnostics. The platform facilitates the integration of wearable sensor data for tailored therapy of sleep disorders.