Prediction of the glucose level in blood using near infrared spectrometer

Abstract

Diabetes is a medical condition caused by high glucose level in blood. A good control of glucose level is essential for diabetes patients as negligence of management could lead to severe health conditions such as obesity, blindness, stroke or heart attack. A conventional glucose level assessment uses an invasive glucometer device that is widely used in clinical practice. However, this practice is unfavourable for some patients since pricking fingers multiple times per day can be tiresome and painful, besides causing calluses, and not preferred by individuals with dexterity limitation, algophobia or anxiety problems. Therefore, prediction system of glucose level using non-invasive method is widely investigated. This thesis presents development of a non-invasive prediction system of glucose level in blood using near infrared spectrometer (NIRs), combined with predictive linear and non-linear models. This research focuses on the capability of spectrum to penetrate the skin, as well as the correlation between skin depth, location of human blood vessel and length of NIR spectrum. The data utilized are acquired from three groups: the existing diabetic patients, the non-diabetic persons, and a group of persons with no prior diagnosis. The existing diabetic patients are under medical treatment from Hospital Universiti Sains Malaysia (HUSM), while the non-diabetic persons are subjects who had their medical check-up in less than one year prior. Meanwhile, the control group consists of subjects who never had their medical check-up in the last 2 years. The glucometer is treated as reference data, and both glucometer and NIR spectral readings were obtained from all subjects (1000nm-2000nm). From the wavelength, regions that show significant information of glucose and water are between 1440 nm – 1460 nm and 1940 nm – 1960 nm. Results from pre-processing stage imply that data pre-processed by Savitzky-Golay (SG) filter with optimal parameter setting achieved the best accuracy. To establish correlation between the reference data and the NIR spectrum, two linear models, Autoregressive with Exogenous (ARX) and Autoregressive Moving Average Exogenous (ARMAX) models were implemented, and the combination of ARX and ARMAX with Artificial Neural Network (ANN) were utilized as non-linear models. The unregularized and regularized models for both ARX and ARMAX show unsatisfying results, where unregularized ARX is only at 24.82%, regularized ARX at 36.40%, unregularized ARMAX at 53.89% and regularized ARMAX shows accuracy of 78.57%. The results from regularized ARX and ARMAX are then used to combine with the ANN models. The ARMAX-ANN result shows a significant improvement at 89.45% respectively. The Clarke Error Grid Analysis (CEG) was used as a method to validate the new system with the reference established method in clinical practice. The CEG analysis reveals that the distribution of samples lies in region A and region B, where region A is within 20% of the reference sensor and region B is outside of 20% but would not lead to an inappropriate treatment for patients. From the results obtained, it is concluded that the selection of NIR regions and non-linear ARMAXANN model is proven as a promising method in predicting the glucose level in blood and future works can be executed in enhancing system accuracy.