Non-Invasive Near-Infrared-Based Optical Glucose Detection System for Accurate Prediction and Multi-Class Classification

Abstract

One of the most common diseases around the world is diabetes. Intrusive methods involving blood samples via a finger prick are required to test for diabetes. These treatments are uncomfortable and prone to infection. Non-invasive testing is proposed as a solution to this concerning problem. To test the glucose levels of subjects, a shortwave near-infrared-based optical detection system with a 950 nm wavelength sensor in reflective mode is presented. The system collects the measured signal through voltage, transmittance, absorbance and reflections to estimate glucose. The relation between voltage and predicted glucose is evaluated from the absorbance, reflectance, and voltage for 575 samples. A Multiple linear regression (MLR) expression is used in the proposed method to enhance the accuracy. The proposed method achieves a coefficient of determination (R2) of 99% and a mean absolute derivative of 3.6 mg/dl in real-time data analysis with the sensor. The root mean square error (RMSE) is also calculated as 3.46 mg/dl. Three additional machine learning classifiers are employed to achieve high accuracy in multi-class classification. Adaboosting and Gaussian Naïve Bayes classifiers achieve an accuracy of 97% each. Furthermore, the system computes performance metrics such as precision, recall, and F1-score, and predicts the class on the test sample.