The utilization of circadian rhythm features to improve sound-based AI sleep staging

Oct 20, 2023
World Sleep 2023

Abstract

Introduction:
Due to the increased interest in sleep and its significance to public health, studies have been conducted to develop simple and easy ways to assess sleep quality using smartphones. One such approach involves utilizing sleep sounds to predict sleep stages. Although sleep sounds provide valuable information concerning respiratory patterns and muscle movements, circadian rhythms can supply additional insights into temporal dynamics and regulatory mechanisms of sleep. Hence, this study aims to enhance sound-based AI sleep diagnosis algorithms by incorporating circadian rhythm features.

Materials and Methods:
We developed a deep neural network model that inputs sleep sounds and circadian rhythm components. This AI model classifies sleep into four stages: Wake, REM, Light, and Deep. Three circadian rhythm features as inputs for the model were derived from the two-process model of sleep regulation: (i) melatonin cycle, (ii) sleep pressure, and (iii) time in sleep. The melatonin cycle was modeled using the cosine function, mimicking the natural melatonin cycle, where it captures the increase at the beginning of sleep, the peak in the middle of the night, and the subsequent decrease to low levels by early morning. The sleep pressure, which is known to exponentially decrease since the onset of sleep, was modeled using an exponential decay function. The time in sleep, which is represented by the elapsed time since the onset of sleep to inform the model of the relative position during whole night, is characterized by a linear function. A pre-trained model called MobileViTv2 was first developed and fine-tuned with a cross-entropy loss function. A labeled hospital dataset (PSG and audio, N=2,574) was used to train and validate the model, and the trained model was evaluated on the labeled data recorded on a smartphone at home (PSG and audio, N=128).

Results:
The final model achieved an accuracy of 70.44% and a Macro-F1 score of 0.6686 for sleep staging in a home environment. In the absence of circadian rhythm features, the model only reached an accuracy of 67.00% and a macro-F1 score of 0.6395. There was a performance improvement of 3.44%p and 2.91%p in accuracy and macro-F1 score, respectively. These results emphasize the substantial contribution of circadian rhythm features in improving sleep stage prediction, even with simple modeling. Furthermore, ablation studies were conducted to examine the individual impact of each circadian rhythm feature. Among the three features, sleep pressure demonstrated the highest accuracy, reaching 69.02%. When only two out of the three features were used with different combinations, the sleep pressure-melatonin cycle pair yielded the highest accuracy of 69.43%. Overall, the model with all three circadian rhythm features exhibited the best results in sleep stage analysis.

Conclusions:
To enhance the stage predicting sleep sound-based AI model, the basic concepts of the circadian rhythm were employed. This easy-to-use and accurate sleep staging model will be the first step to help improve sleep-related issues, leading to better overall health outcomes.

Authors

S. Park
E. Cho
H. Park
J. Jung
D. Lee
D. Kim
J.-W. Kim
K.-Y. Jung

Acknowledgments

1. I confirm that the abstract and that all information is correct.: Yes
2. I confirm that the abstract constitutes consent to publication.: Yes
3. I confirm that I submit this abstract on behalf of all authors.: Yes
I understand that the presenting author MUST register for the congress.: Yes