JMIR Publications

ABSTRACT

Background:

With a growing interest in sleep monitoring at home, sound-based sleep staging using deep learning has emerged and been validated using in-laboratory sounds. However, validation in noisy home environments has not been carried out despite its importance.

Objective:

To develop and validate a deep learning method to perform sleep staging using smartphone audio recordings in uncontrolled home environments.

Methods:

The training of the model consists of three components: (i) supervised learning using 812 pairs of in-laboratory polysomnography (PSG) and audio recordings, (ii) transfer learning from hospital to home sounds by adding 829 smartphone audio recordings at home, and (iii) consistency training using augmented in-laboratory sound data. Augmented data were created by adding 8,255 home noise data to in-laboratory audio recordings. An independent dataset was built by collecting matched level 2 PSG and smartphone audio recordings at home from 45 individuals to examine the performance of the trained model.

Results:

The accuracy of the model was 76.2% (63.4% for wake, 64.9% for rapid-eye movement (REM), and 83.6% for non-REM). The Macro F1 score and mean per-class sensitivity were 0.714 and 0.706 respectively. The performance was robust across demographic groups such as age, gender, body mass index, or sleep apnea severity (accuracy 73.4-79.4%). Transfer learning and consistency training enhanced the performance by 7% of accuracy.

Conclusions:

This study shows that sound-based sleep staging using smartphones in noisy home environments is feasible. People may easily monitor their sleep at home using their own smartphones without an additional device. Clinical Trial: N/A