Bringing hearables to life: applications in sleep and glucose monitoring
File(s)
Author(s)
Hammour, Ghena
Type
Thesis
Abstract
This thesis explores an innovative use of hearables—in-ear sensing of neural function and vital signs—related to monitoring sleep and blood glucose levels. Through analysis and case studies, we demonstrate the utility and potential of in-ear electroencephalogram (EEG) and in-ear photoplethysmogram (PPG) devices for sleep staging and glucose monitoring. We first introduce a non-invasive method for automatic sleep staging in older adults using a single-channel ear-EEG approach. By extracting features from frequency, time, and structural complexity domains, substantial agreement with gold-standard human-scored hypnograms was achieved, with a kappa value of at least 0.61. This offers a portable and cost-effective alternative to traditional polysomnography.
The second case study leverages existing scalp EEG data to enhance in-ear EEG analysis through transfer learning. By fine-tuning scalp-EEG pre-trained models using ear-EEG data, classification accuracy improved by 4%, paving the way for a transition from standard wearables to hearables.
In physiological sensing, motion artifacts are often seen as nuisances. This work takes the approach that “no data is bad data” by using artifacts from hearables as behavioural cues. Activities such as sitting, speaking, chewing, and walking were classified from real-world data. Analysis using various machine learning techniques yielded high accuracy and introduced a novel approach to human activity recognition. Finally, we address a novel application in non-invasive glucose monitoring using an in-ear PPG device. This work establishes a relationship between raw PPG and blood glucose levels, with 82% of estimates falling within clinically acceptable Clarke error grid regions. Overall, this thesis demonstrates the promise of hearables for unobtrusive, continuous, and accessible health monitoring.
The second case study leverages existing scalp EEG data to enhance in-ear EEG analysis through transfer learning. By fine-tuning scalp-EEG pre-trained models using ear-EEG data, classification accuracy improved by 4%, paving the way for a transition from standard wearables to hearables.
In physiological sensing, motion artifacts are often seen as nuisances. This work takes the approach that “no data is bad data” by using artifacts from hearables as behavioural cues. Activities such as sitting, speaking, chewing, and walking were classified from real-world data. Analysis using various machine learning techniques yielded high accuracy and introduced a novel approach to human activity recognition. Finally, we address a novel application in non-invasive glucose monitoring using an in-ear PPG device. This work establishes a relationship between raw PPG and blood glucose levels, with 82% of estimates falling within clinically acceptable Clarke error grid regions. Overall, this thesis demonstrates the promise of hearables for unobtrusive, continuous, and accessible health monitoring.
Version
Open Access
Date Issued
2024-07-10
Date Awarded
01/10/2025
License URL
Advisor
Mandic, Danilo
Publisher Department
Department of Electrical and Electronic Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)