The electroencephalography (EEG) is a standard for studying the dynamics of brain activity, especially that related to cognition and disease. However, despite the EEG being extensively employed in both clinical settings and brain-computer interface (BCI), real world EEG-based applications are not yet well established. This is largely due to the cumbersome and obtrusive nature of current EEG sensing which has been prohibitive to the continuous acquisition of data over sufficiently long periods of time. To bring the power of EEG monitoring of neural activity out of clinic, this thesis explores the feasibility of EEG monitoring in the real world using the recent discovery of ear-canal based physiological sensing, termed in-ear EEG. The utility of wearable in-ear sensing technology, especially its capability to monitor brain activity for automatic sleep monitoring and person authentication, is investigated.
This thesis first demonstrates the utility of in-ear sensing for one of the cornerstones of well-being in modern society – sleep monitoring. Sleep is an essential process for human well-being, and its quality reflects both a person’s lifestyle as well as various medical conditions. Current clinical sleep monitoring analysis reflects sleep profiles of individuals rigorously and comprehensively, however, the cumbersome nature of standard recordings can disturb patients’ normal sleep, while the scoring process places enormous demands on a clinician’s time. To mitigate these issues, feature extraction for automatic sleep monitoring is investigated. The combination of structural complexity analysis and non-linear frequency domain features is shown to be effective for classifying sleep stages, and the feature-fusion based classification model using those features is shown to outperform other existing studies using a publicly available overnight sleep EEG dataset. In addition, the efficacy of smoothing algorithms implemented by a hidden Markov model is investigated in this context. Following the investigation of automatic sleep staging algorithms, the feasibility of an in-ear EEG device for large scale sleep monitoring is scrutinised. To begin with, as proof-of-concepts, numerous nap recordings are conducted inside the laboratory with in-ear EEG. To provide wearable sleep monitoring in real-world, simultaneous conventional polysomnography (PSG) and in-ear EEG during the night is recorded in subjects’ home; the agreement between automatic scoring based on in-ear EEG and manual scoring based on the PSG was 74.1% and the corresponding kappa coefficients was 0.61. Through numerous recordings and analysis, a wearable in-ear EEG device for sleep monitoring is proposed and validated as a 24/7 continuous and unobtrusive technology for sleep quality assessment in the community.
The other major in-ear EEG application explored is person authentication, which refers to the process of confirming the claimed identity of an individual. The use of EEG as a biometrics modality has been investigated for the last 15 years; however, its feasibility in real world has not yet been established, mainly due to issues of reproducibility. To this end, a readily deployable EEG biometrics system based on in-ear EEG is proposed, that is user-friendly and discreet in nature and does not require supervision by skilled personnel or the use of cumbersome equipment. For rigour, recordings over two days from 15 subjects are conducted; both frequency domain and autoregressive features are employed to identify an individual. The achieved classification accuracy was 95.7% in the verification setup whereas the kappa coefficient was 0.65 in the identification setup.
Overall, the comprehensive analysis over a number of recordings and proof-of-concept studies in this thesis demonstrates the feasibility of the proposed in-ear EEG based applications. The so-verified wearable in-ear EEG paradigm promises faithful EEG applications in real-world environments, and resolves the critical issues with robustness associated with current use of EEG.