Non-Invasive In-Vitro Glucose Monitoring Using Optical Sensor And Machine Learning Techniques For Diabetes Applications

Diabetes is a major public health challenge affecting more than 451 million people. Physiological and experimental factors influence the accuracy of non-invasive glucose monitoring, and these need to be addressed before replacing the finger prick method with a non-invasive glucose measurement technique. Also, the suitable employment of machine learning techniques on experimental data can significantly improve the accuracy of glucose predictions. This work includes the design, development, testing and data analysis of an optical based sensor for glucose measurements. The feasibility of non-invasive measurement of glucose within aqueous solutions that assimilate the composition of human blood plasma is investigated. The laboratory testing of the sensor with controlled solutions helps to make valid conclusions about the performance of the sensor and the accuracy of the predictions. There are several goals associated with this study. The first goal is to use light sources with multiple wavelengths to enhance the sensitivity and selectivity of glucose detection in the aqueous solution. Multiple wavelength measurements have the potential to compensate for errors associated with inter- and intra-individual differences in blood and tissue components. In this study, the transmission measurements of the custom built optical sensor are examined using 18 different wavelengths between 410 and 940 nm. Results show a high correlation value (0.98) between glucose concentration and transmission intensity for 4 of the 18 wavelengths (485, 645 and 860 and 940 nm). The second goal of this study is to extract the intensity data using these 4 wavelengths and to collectively analyze the accuracy of glucose concentration predictions based on machine learning techniques. The intensity data measured using the four optimal wavelengths is first input into a support vector machine (SVM) classifier to discriminate hypoglycemic, normal and hyperglycemic ranges with 99% accuracy based on F1-scores. Then, for each of the three ranges, a feed-forward neural network model is developed and applied to a test dataset to predict the glucose concentration within each range separately. The combination of both the classification and regression models to predict glucose concentrations for a given dataset results in more accurate and reliable predictions compared to using a single regression model for the entire glucose range. The use of hybrid models improves the root mean square error (RMSE) from 12.7 mg/dL (in the case of a single regression model) to 9.3 mg/dL (in the case of hybrid models as was done in this study). Using this hybrid approach also results in 100% of the glucose readings falling within zones A and B of the Clarke error grid. This is an important step towards critical diagnosis during an emergency patient situation. Future work should include an in-vivo study to test the sensor and data analysis approach used here on human blood samples