Polysomnography from the inner ear for consumer and clinical use

General Area:
Polysomnography (PSG) is the measurement of several physiological signals including electroencephalography (EEG), electrocardiography (ECG), electromyography (EMG), respiratory flow and more, during sleep to assess sleep health. The importance of a portable non-restrictive PSG recording device becomes clear when current clinical applications are considered. For example, to assess sleep health a patient must stay in an overnight clinic and wear the EEG skull cap, heart rate monitor, respiration tubes and many other sensors. This is time consuming, expensive, and has the potential to severely bias results by making the sleep uncomfortable. A portable device would solve these problems and could be implemented by the patient themselves.

In recent years there has been an introduction of a new recording method, the ear-EEG, which has been shown to be as effective as scalp electrode EEG for characterising sleep. Single ear ear-EEG is simple, consisting of just 3 electrodes (ground, reference, and main), positioned on the lobule, helix and within the ear canal. Thus, ear-EEG presents the advantage of being far less restrictive, is implementable by a patient or consumer and provides a platform for portable recording.

Research Direction:
The proposed research will focus on the development of novel machine learning and signal processing solutions for assessment of sleep health such as automatic sleep staging algorithms, in conjunction with further development of the ear-EEG hardware to add additional sensors for PSG and enable classification in an online fashion. The developed solutions will then be implemented in sleep and drowsiness studies.

Research timeline:
The proposed research project is best divided into 3 main sections, algorithms, hardware, and combined application. Algorithm development will occupy the majority of year 1, as will be continued into year 2 and 3. Hardware implementation will occupy the majority of year 2. The application of both machine learning algorithms and ear-PSG hardware to problems such as sleep health, sleep onset detection in drivers and brain injury will begin midway through year 2 and occupy the majority of year 3.

Algorithms:
For automatic classification of sleep stages or sleep onset to be clinically relevant, it must have a high accuracy when compared the ground truth of a clinician. Many current deep learning implementations achieve sufficiently high accuracy but are too computationally complex to be used in a small device performing classification in an online fashion. The first task is to develop simpler algorithms which rely on the transition time series structure of sleep, which achieve both high accuracy and computational simplicity. Novel signal processing algorithms such as panorama and delay vector variance will also be expanded on for applications in sleep staging.
Hardware and application:
The integration of the current Ear-EEG device with different physiological recordings and wireless data transfer will allow for automatic online scoring of sleep stages and drowsiness. This will open the door for large-scale long-term sleep studies which up to this point in time have not been possible. The success of this project will lead to a major advancement in large scale sleep data acquisition and pave the way for a better understanding of sleep patterns and their influence on health.
Current Work:
Classifiers have been implemented using the Cleveland Children's health study data set from sleepdata.org. Current results of the transition probability model are very promising, with an increase in average N1 accuracy of 15% and 25% upon random forest and linear discriminant analysis classifiers respectfully.
EPSRC aligned research areas:
Digital Signal Processing, Microelectronic device technology, Artificial intelligence technologies, Mathematical Biology