Objective:
There is a great interest in observing breathing patterns during sleep, as sleep disturbances can be caused by respiratory irregularity and cessations. In this paper, we introduce the first steps to an accelerometer-based screening tool for respiratory rate estimation and a novel approach towards detecting breathing cessations such as apnea/hypopnea, by extending and combining established signal processing routines with machine learning.

Methods:
From a single chest-worn accelerometer, we estimate the respiratory rate based on the inhalation/exhalation movements of the chest and carry out a full overnight validation. On this basis, we build a set of features customized to detect irregular respiratory activity, including a novel feature: the respiratory peak variance (RPV). From thirteen healthy subjects, a classification model was trained, validated, and tested with over 98 h of PSG-labeled accelerometer data.

Results:
The algorithm estimated the respiratory rate with a mean difference of 1.8 breaths per minute compared to respiratory inductance plethysmography during overnight PSGs. The machine learning classifier detected respiratory cessations with a sensitivity and specificity of 76.05% and 70.05% respectively, with an overall accuracy of 70.95%.

Conclusion:
We successfully demonstrated the potential of a novel respiratory feature set in a preliminary application with young healthy volunteers for respiratory rate estimation and in identifying apnea/hypopnea events during overnight sleep.

Significance:
We present a simple and unobtrusive wearable system that can serve as a home screening tool for sleep-related breathing disorders.