Noninvasive Glucose Detection Using Infrared Photoacoustic Spectroscopy and Machine Learning

Abstract

The ideal method to monitor diabetes is to obtain the glucose level with a fast, accurate, and pain-free measurement that does not require blood drawing or finger pricking. Although the development of noninvasive devices for blood glucose measurement started three decades ago, no clinically proven devices were commercially released in the market. Among all the noninvasive glucose detection techniques, optical spectroscopy has rapidly advanced, including \gls{ir} and \gls{Pa} spectroscopy.

The combination of \gls{ir} and \gls{Pa} spectroscopy has shown promising developments in recent years as a substitute for invasive glucose monitoring technology. The \gls{ir} region has a strong relationship with glucose due to the presence of glucose absorption peaks. \gls{Pa} spectroscopy utilizes the vibration modes of the glucose molecules in the \gls{ir} region and the weak water absorption of acoustic signals as an alternative approach to compensate for the optical losses in the \gls{ir} transmission and absorbance spectroscopy. The concept of \gls{Pa} spectroscopy relies on generating acoustic waves, by an electromagnetic source, that are distinguishable from one material to another and can be detected by sensitive ultrasonic or piezoelectric sensors.

The first part of the thesis demonstrates the development of the \gls{ir} and \gls{Pa} system for noninvasive glucose monitoring. The \gls{ir} and \gls{Pa} system has been developed using a single wavelength \gls{qcl}, lasing at a glucose fingerprint of 1080 cm$^{-1}$. In biomedical applications, phantoms are widely used as test models to substitute targeted body objects. Biomedical skin phantoms with similar properties to human skin have been prepared at different glucose concentrations of $\pm$25 $mg/dL$ as test models for the setup. The system shows feasibility in detecting glucose using these skin phantoms, covering the normal and hyperglycemia blood glucose ranges. Machine learning classification models have been employed to enhance the prediction accuracy of glucose levels using unprocessed acoustic signals.

The second part of the thesis extends the development of the \gls{ir} and \gls{Pa} system. A dual single-wavelength \gls{qcls} system has been developed using \gls{Pa} spectroscopy for noninvasive glucose monitoring. The glucose detection sensitivity of the \gls{ir} and \gls{Pa} spectroscopy has improved to $\pm$12.5 $mg/dL$ using dual \gls{qcls} lasing at 1080 \& 970 $cm^{-1}$. The artificial skin phantoms have been prepared with other blood components at different glucose concentrations. The dual \gls{qcls} system demonstrates sustainability in detecting glucose concentrations in the presence of albumin, sodium lactate, cholesterol, and urea. An ensemble classifier model has been developed to predict the glucose level of skin samples. The model has achieved 96.7\% prediction accuracy for samples with and without blood components, with 100\% of the predicted data located in zones A and B of Clarke's \gls{ega}.

After demonstrating the glucose detectability of the \gls{ir} and \gls{Pa} system for the \textit{in vitro} measurements, the system has progressed to the \textit{in vivo} experiments. The operating power of the \gls{qcls} has been lowered to fulfill the safety guidelines of using light sources on human skin. The blood glucose concentration can be potentially measured from the \gls{isf} located underneath the skin in the epidermis layer. The glucose diffuses from the blood to the \gls{isf} layer, creating a significant opportunity for noninvasive monitoring systems. The \textit{in vivo} measurements of the fiber-coupled dual \gls{qcls} and \gls{Pa} system have been assessed by oral glucose tolerance test \gls{ogtt}. The preliminary results from the \textit{In vivo} measurements demonstrate that the \gls{mir} and \gls{Pa} system can detect glucose levels not only in the biological samples but also in real human skin. Finally, a Gaussian Process regression model has been developed to improve the prediction accuracy of the \gls{ir} and \gls{Pa} system.