Design Engineering
Showcase 2020

With only one third of adults across developed nations obtaining the recommended eight hours of sleep, insufficient sleep quality is an unrecognised problem that has adverse consequences in a huge variety of aspects [1].

Sleep is influenced by fixed parameters such as gender, age and genetics, by daily contextual factors such as exercise level, sleep routine and alcohol intake that require behaviour change to alter, and by the bedroom microclimate including light, temperature, humidity and sound level [1][2].

Given the rising trend in automated smart home technologies, the opportunity for night-time microclimate actuation seems a very feasible route to improving sleep and is where this research proposal stems from [3].

1. Can a system be built to identify microclimate-induced abnormal sleep behaviour in real time based on physiological, contextual and local microclimate data monitored by existing consumer smart home and sleep monitoring technologies?
2. If this proves viable, is it possible to derive the cause of the abnormal events so that relevant smart home technologies may be integrated into the system, to adapt the microclimate in real time and prevent the abnormal event from occurring?

Before answering this proposal, it’s important to understand how to asses sleep quality. Sleep quality is reflected in sleep composition, formed by several cycles that comprise of distinct stages. During NREM (non-rapid eye movement) stages, heart rate, respiratory rate and brainwave activity all slow down. Whereas, during REM (rapid eye movement), where dreaming typically occurs, these biometrics all increase [1]. Therefore, by analysing biometric features it’s possible to classify sleep stage and assess sleep quality, which is what’s done in sleep science and most consumer sleep assessment technologies.

A method was devised to address the research proposal:

Two predictive models can be built, one to classify current sleep (Model 1) and another to predict sleep in an ideal microclimate in a specific contextual circumstance (Model 2). Discerning between the two classifications will identify microclimate-induced abnormal sleep behaviour. By cross-referencing the timestamp of abnormal behaviour with microclimate parameters, the cause can be identified and then rectified with microclimate actuation devices. Additionally, to prevent abnormal sleep behaviour ahead of it occurring, sleep stage transition classification is necessary, in addition to sleep stage classification.

Monitoring

The Emfit QS+Active under-mattress device was chosen to monitor sleep, it detects a ‘ballistocardiograph’ (or ‘BCG’) signal, a measurement of force. To monitor the bedroom microclimate, a Raspberry Pi was setup with various microclimate sensors. Additionally, a diary was kept to record relevant daily contextual data.

Signal Processing

The BCG signal shows large- and small-scale fluctuations representing body movement and heart/respiratory pulsations respectively. Heart rate, heart rate variability, activity and activity level variability were all calculated by extracting relevant peaks & waveforms from the signal.

Feature Selection

Feature selection was used to identify features relevant to the labelled data. Data was labelled with four sleep stages and the twelve possible transition variations with transitions labelled as the 30 second period preceding a change of sleep stage. Feature selection was conducted with biometric features extracted from the BCG signal for Model 1’s input and sleep composition features extracted from Model 1’s output for Model 2’s input.

Machine Learning

Recurrent Neural Networks were used, a class of Neural Networks that recognise patterns in sequential data. The models were run with 20 nights of data split into a training, validation and test set. Models 1 & 2 achieved accuracies of 49.3% & 49.4% on the test sets respectively. It’s important to note that to build accurate models much more data is necessary, especially for classifying transitions that form such a small portion of the data.

Using insights from this process, a system was proposed to improve sleep quality via optimisation of the local microclimate.

The bedroom microclimate is monitored by various smart home devices. Sleep quality is monitored by an under-mattress BCG sensor from which biometric features are extracted, used as input for Model 1 for real time sleep stage/transition classification. Daily contextual factors are monitored by calendar data, wearables etc. The system builds a selection of predictive models from nights categorised by contextual data. The contextual data from this day is used to select the model suitable for this night: Model 2. This model bases sleep stage/transition classification on sleep composition in an ideal microclimate given a particular contextual circumstance. It takes sleep composition feature input from Model 1’s output. If a transition is predicted, predictions are compared and if discrepancies are above a set threshold, this indicates a microclimate induced abnormal transition and the microclimate is checked for fluctuations that indicate the cause. Once communicated to relevant smart home devices, the abnormal transition can be stopped from proceeding via corrective actuation.

1. Walker M. Why We Sleep: The New Science of Sleep and Dreams, 1st ed. London, UK: Penguin; 2018.

2. National Sleep Foundation, Sleepfoundation.org, 2020, Available from: https://www.sleepfoundation.or...

3. Almost a quarter of Britons now own one or more smart home devices | YouGov, Yougov.co.uk. 2020, Available from: https://yougov.co.uk/topics/te...