Machine learning and pervasive sensing for sleep assessment and health

A good night's sleep is the foundation for a healthy life. The insufficient long-term sleep can result in a various negative influence on our lives, such as daytime drowsiness and impaired cognitive function, heart disease, high blood pressure and type II diabetes.
Many existing published papers focused on the end-to-end machine learning solutions that consist of automatic sleep-wake classification, sleep disorder assessment, etc. Such models intake the raw signal data such as the Polysomnography(PSG) or Actigraphy (similar consumer product like Fitbit, Jawbone etc.) as the inputs to generate the final predictions such as to distinguish sleep or wake status. Most of those studies were completed based on a single modality data source such as a few channels of EEG. The multimodality fusion is one of the appealing methods to improve the prediction and classification accuracy such as to include heart rate, respiration and limb movements' accelerometer data.

In the past decade, open source and commercial wearables, Body sensor network (BSN), ambient and IoT technologies have been well used in sleep and daily activity research. But accurately monitoring sleep-wake pattern or even the sleep stages over a long period is still a challenging task in a free-living environment. Since human behaviour consists of complex activities and the environmental factors are non-static, it is difficult for a single modality sensor data to capture all activities accurately. For instance, the WIFI and Radar based method works well on a single person sleep monitoring scenario but they are unable to distinguish a couple sleep situation. Multiple heterogeneous sensing methods can improve Signal-to-Noise Ratio, extend parameter coverage and integrate the independent features and prior knowledge.

Aims
To understand how daily activity pattern, lifestyle and comorbidities affect people's sleep parameters, stages and subjective feelings by utilising the multimodality data fusion and transfer learning method based on the public dataset (e.g. UK biobank, White Hall II).
To investigate the adoption of pervasive sensing technology (e.g. IoT devices) and the developed multimodality framework to encourage behaviour change in public.

Methodology
The first stage of the PhD work will focus on the development of algorithm, models and frameworks based on existing sleep and health dataset that includes Fenland II Dataset, UK Biobank, and Whitehall II dataset by adopting the numerical analysis mathematical modelling. The study will consider the published work from computer vision and machine learning as the starting point.

The second stage of the PhD work will focus on the adoption of pervasive sensing technology and machine learning method to develop an inexpensive wearable and nearable system to better monitoring sleep quality and duration. The study will collect multi-modality data from sensors (audio, movement, temperature, radar, etc.) and the ground truth PSG sleep data through the recruited participants. Afterwards, the first stage developed framework will be used to investigate the relationship between physical activities, sleep, lifestyle and non-communicable diseases

The developed system will provide opportunities for researchers, clinicians and end users to understand how those factors will impact their health so as to encourage sleep behaviour and the change of lifestyle.