Contactless Sleep Staging With Radar: A Transfer Learning Approach.

Area of Science:

• Biomedical Engineering
• Sleep Medicine
• Artificial Intelligence

Background:

• Accurate sleep monitoring is crucial for diagnosing sleep disorders but traditional methods are costly and impractical for long-term use.
• Contactless radar sensing offers a non-intrusive alternative for real-world sleep tracking.
• Limited availability of large, labeled datasets hinders the development of robust radar-based sleep classification models.

Purpose of the Study:

• To improve sleep stage classification accuracy and generalization using transfer learning with radar data.
• To address the challenge of limited labeled datasets in radar-based sleep monitoring.
• To evaluate the effectiveness of pretraining on a large dataset (MESA) for fine-tuning a radar sleep analysis model.

Main Methods:

• Utilized transfer learning by pretraining an LSTM model on movement, heart rate variability (HRV), and respiratory features from the MESA Sleep dataset.
• Fine-tuned the pretrained LSTM model using radar data from 44 synchronized polysomnography recordings.
• Assessed classification performance using the Matthews Correlation Coefficient (MCC) for five-class sleep staging.

Main Results:

• Transfer learning significantly improved the MCC from 0.25 to 0.47 for five-class sleep staging.
• Accuracy gains were particularly notable for Wake, N3, and REM sleep stages.
• The approach demonstrated enhanced generalization to unseen participants within the radar cohort.

Conclusions:

• Radar-based sleep analysis, enhanced by transfer learning, shows significant potential for scalable, contactless, long-term sleep monitoring.
• The study highlights the feasibility of leveraging large existing datasets to overcome data limitations in new sensing modalities.
• Future research should focus on cross-modal domain adaptation to further refine radar sleep analysis.