A Flexible Deep Learning Architecture for Temporal Sleep Stage Classification Using Accelerometry and...

Area of Science:

• Biomedical Engineering
• Sleep Science
• Artificial Intelligence

Background:

• Wrist-worn consumer sleep technologies (CST) with accelerometers (ACC) and photoplethysmography (PPG) are prevalent for out-of-clinic (OOC) sleep monitoring.
• Limited validation studies exist due to restricted access to raw CST data.

Purpose of the Study:

• To develop and validate a deep neural network (DNN) for predicting sleep stages using ACC and PPG signals from CSTs.
• To assess the generalizability and robustness of the DNN on internal and external datasets, including raw data from wrist-worn devices.

Main Methods:

• A U-Net inspired DNN was designed to process multivariate time series data (ACC and PPG) for sleep stage prediction.
• The DNN was trained and validated on internal datasets (301 recordings) and externally validated on a hold-out test set (35 recordings) of wrist-worn CST data.
• Spectral preprocessing was compared against surrogate, feature, and raw data preparation methods.

Main Results:

• The DNN achieved accuracies of 0.71-0.76 and κ values of 0.58-0.64 on internal test sets.
• Spectral preprocessing outperformed other data preparation methods.
• Combining ACC and PPG modalities yielded the best performance, with PPG crucial for REM sleep detection and ACC improving wake/sleep estimation.
• External validation on wrist-worn CST data resulted in an accuracy of 0.69 and κ of 0.58, demonstrating generalizability.

Conclusions:

• The developed DNN demonstrates a robust and generalizable approach for OOC sleep stage prediction using wrist-worn CST data.
• Spectral preprocessing and multimodal data fusion (ACC and PPG) significantly enhance prediction accuracy.
• This work supports the potential of CSTs as validated tools for sleep monitoring outside clinical settings.