Attention-Based LSTM for Non-Contact Sleep Stage Classification Using IR-UWB Radar

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Sleep Medicine

Background:

• Polysomnography (PSG) is the gold standard for sleep staging but requires specialized personnel and uncomfortable sensors.
• Non-contact sleep staging methods using machine learning are emerging as a less invasive alternative.
• Remote vital sign detection offers a promising avenue for unobtrusive sleep monitoring.

Purpose of the Study:

• To develop and evaluate an attention-based bidirectional long short-term memory (Attention Bi-LSTM) model for automatic sleep stage scoring.
• To utilize impulse-radio ultra-wideband (IR-UWB) radar for non-contact remote detection of vital signs for sleep analysis.
• To compare the performance of the proposed Attention Bi-LSTM model against conventional LSTM networks.

Main Methods:

• Sixty-five healthy volunteers underwent simultaneous nocturnal PSG and IR-UWB radar measurements.
• Features related to movement, respiration, and heart rate variability were extracted from IR-UWB signals.
• An Attention Bi-LSTM model was trained, validated, and tested on sleep stage classification using extracted features.

Main Results:

• The Attention Bi-LSTM model achieved an accuracy of 82.6 ± 6.7% and a Cohen's kappa of 0.73 ± 0.11 for classifying wake, REM, light, and deep sleep stages.
• Performance was significantly higher than conventional LSTM networks (p < 0.01).
• Results surpassed those reported in comparative studies, highlighting the model's effectiveness.

Conclusions:

• The Attention Bi-LSTM model demonstrates significant effectiveness for automatic sleep staging using non-contact radar-based vital sign monitoring.
• The integration of attention mechanisms with Bi-LSTM networks enhances sleep stage classification accuracy.
• This non-contact approach offers a promising alternative for sleep quality and architecture assessment.