Comparison of signal combinations for cardiorespiratory sleep staging

Miriam Goldammer1, Sebastian Zaunseder2, Franz Ehrlich1, Hagen Malberg1
1TU Dresden, 2FH Dortmund


Abstract

Introduction: Sleep staging from cardiorespiratory signals has improved significantly during the last two decades. Increasingly, feature-based classifiers are replaced by approaches that use signals or time-series as inputs to neural networks. So far, single signals (mainly electrocardiogram) have been the main focus of investigations, even though we know from feature-based approaches that information from respiration and cardiorespiratory coupling will result in significant classification improvements.

Methods: We modified our previous neural network model, consisting of convolutional and recurrent layers, to take different signal combinations as input. We added oxygen saturation and different respiratory signals to the electrocardiogram, which we considered beneficial according to literature. We further invoked different processing strategies that have been described previously for such signals, namely using downsampled signals vs. using time-series of breath-to-breath intervals. We used data from polysomnograms of the Sleep Heart Health Study and split them into training (3867 subjects) and hold-out test (916 subjects) data. We extracted electrocardiograms, thoracal and abdominal respiratory inductance plethysmograms and sleep stage labels from each polysomnogram. The labels were transformed to match the five sleep stages according to the American Academy of Sleep Medicine. After preprocessing, the signals became input for our convolutional recurrent neural network architecture. Mean Cohen’s kappa κ served as key metric for evaluation.

Results: We found the best combination of signals to be electrocardiogram together with both downsampled respiratory signals and oxygen saturation. The classification resulted in a κ of 0.69 on hold-out test data, which outperforms our previous results and state of the art for cardiorespiratory sleep staging.

Conclusion: We observe that combinations of cardiorespiratory signals can further improve classification performance for automatic cardiorespiratory sleep staging. As there are generally more cardiorespiratory signals available and many more options for preprocessing them, we expect that further research in this area will show even more improvements.