| dc.description.abstract | Diabetes is a chronic condition that occurs when the levels of glucose are high in the blood because the body cannot produce any or enough insulin or use insulin effectively. According to the International Diabetes Federation (IDF), 537 million people are currently living with diabetes. It is very important for people with diabetes to regularly check their blood glucose levels to keep track of any increase or decrease in these levels, and adjust the amount of medication accordingly. This process is called Continuous Glucose Monitoring (CGM). CGM techniques can be categorized based on invasiveness as invasive, minimally invasive, and non-invasive. In the non-invasive there is no need for any blood sample extraction or any implantation of electrodes in the body. First, a review of the non-invasive CGM techniques in the last ten years is conducted in order to understand the current status of the CGM and highlight the challenges that face every technique in order to come up with a better solution. The techniques used for non-invasive CGM can be classified into six major categories: optical, microwave, thermal, transdermal, hybrid and other. In order to overcome the shortcomings of the invasive and minimally-invasive methods of CGM, such as pain, discomfort, and risk of infection, non-invasive CGM is needed. However, due to the multiple challenges such as accuracy, usability and applicability, contemporary non-invasive glucose monitors are still not sufficiently reliable.
In this thesis, a non-invasive glucose monitoring system is developed using microwave sensor with machine learning techniques. The system has two parts: hardware, which is the microwave sensor, and software, which is the machine learning algorithms. The physical sensor is microwaves-based using inexpensive printed circuit board technology. Electrically-small dipole and another spiral microwave sensor were designed and used in this thesis taking into account different factors like frequency range, penetration and safety of the human. Machine learning techniques were used to select the most distinguish features in order to predict the actual glucose level in the human. Different feature engineering types were used to extract the discriminate features that will be inputted to different regression algorithms to predict the glucose levels.
The main idea of the thesis is based on studying dielectric properties (permittivity and conductivity) of the human body tissues in order to find a relation with the corresponding glucose level in those tissues. This is done using CST simulation along with experiments.
Experimental results on aqueous solutions (water-glucose solutions) used as a proof of concept and to check the ability of the microwave sensors to detect the different concentrations of these simple water- glucose solutions.
In simulation, a hand model system was designed with different tissues/layers to simulate the effects of the microwave sensor with respect to changing in dielectric properties (permittivity and conductivity) of those tissues/layers. Different systems (corresponding to different hand layers/tissues) were trained and tested using cross validation, and the Root Mean Square Error (RMSE) were acceptable. | en |
