Sensing and signal processing for non-invasive blood glucose monitoring

Remote monitoring is required in several applications and generically, a remote monitoring architecture can be separated into 3 distinct but inter-related layers namely: sensing, signal processing and communications. However, this study focuses on the sensing and signal processing aspects in healthcare. In particular, the research is to investigate sensing and signal processing techniques for a non-invasive approach to type-1 diabetic patients monitoring. Diabetes mellitus is a long-lasting disease and the number of people with diabetes is increasing rapidly worldwide. Managing this disease requires continuous monitoring of blood glucose levels. It is stipulated that avoiding the traditional finger prick method could help improve the adherence and overall management. The present research is concerned with using Fourier transform near infrared spectrometer for the non-invasive measurement of the blood glucose levels. This research has focused on the signal processing and data analysis aspects where a near infrared spectrophotometer has been employed for the sensing to collect practical representative test data. In the signal processing aspects, most of the researchers to date have tended to employ linear regression techniques are the Partial Least Squares Regression (PLSR) and the Principal Component Regression (PCR) based methods and their variants. However, these methods have limitations in practice and have not been translated into a clinical tool. In this project, we target to overcome the current drawbacks of these techniques and in particular their inability to detect the components with low variance by investigating the potential of certain non-linear regression techniques; one of the promising techniques proposed in this research is based on combining a Local Linear Embedded Regression (LLER) with pre-processing. The coupling of bandpass filtering with the novel LLER has been shown to achieve better prediction results than existing methods. A novel regression model called improved support vector regression coupled with Fourier self-deconvolution is also proposed and compared with the linear calibration models under the same conditions. The other proposed model is Partial Least Squares Regression coupled with Frequency self-deconvoluted ReliefF (FSDR-PLSR) which is based on the variance adjustment according to the importance of the features. Finally, two novel pre-processing methods are introduced in this work; i) pre-processing based on Hilbert haung transformation and ii) pre-treatment technique based on coupling digital bandpass filtering with scatter correction techniques.