ALGORITHM FOR ESTIMATING BLOOD GLUCOSE LEVEL FROM PHOTOPLETHYSMOGRAM SIGNAL READINGS BY NEAR INFRARED LIGHT ABSORPTION SPECTROSCOPY­ BASED SENSOR S

• Main Author: Kushirayati, Syifa
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/78140
• Institution: Institut Teknologi Bandung

The use of a glucometer to monitor capillary blood glucose levels involves an invasive method and generates medical waste from disposable strips and needles. Non-invasive methods for estimating blood glucose levels are needed to reduce these impacts. Currently, several research studies related to non-invasive blood glucose estimation methods have met accuracy standards adapted from ISO 15197:2013. However, they require personal calibration and mostly use black­ box algorithms, without considering the fundamental principles of how the sensors work. In this study, testing was conducted on independent variables derived from both modified Beer-Lambert equation for blood glucose estimation. These independent variables were used as inputs for linear regression and weighted KNN ( K-Nearest Neighbors) regression . The baseline values of the independent variables were obtained from PPG signal features at peak wavelengths of 1460 nm and 1650 nm. The 1460 nm wavelength was chosen because water, the most dominant component in blood, has higher absorbance than other biological components. Meanwhile, at the 1650 nm wavelength, glucose has higher absorbance compared to other dominant biological components such as water. Both PPG signals in this study were obtained from voltage readings by a near-infrared spectroscopy-based sensor. Validation data testing in Python software showed that the algorithm using the weighted KNN regression method with independent variables derived from the modified Beer­ Lambert equation produced the best accuracy. The obtained Mean Absolute Relative Difference (MARD) was 0.091, Mean Absolute Error (MAE) was 10.75 mg/dl, and 100% of the results were in zones A and Bin the Clarke Error Grid Analysis with a total of 24 validation data points. In testing data on the ESP32 ,nicrocontroller, a MARD of0.1, MAE of 10 mgldl, and JOO% in zones A and Bin the Clarke Error Grid Analysis were obtained with a total of 20 testing subjects.