Diabetes Mellitus is a disease characterized by inadequate control of blood glucose levels, and it ranks among the leading causes of human mortality, the measurement of which, requires fingertip pricking. Nowadays, non-invasive healthcare monitoring systems based on wearable sensors and machine learning models hold the future of smart health. However, the nature of these methods is such that they get affected by internal physiological and external parameters, which is one of the obstacles to this path. This study aims to introduce a non-invasive blood glucose level classification method based on machine learning analysis. The device’s sensor consists of an optical finger sensor that transmits near-infrared light, filtering, and amplification to reduce the noise of the extracted photoplethysmography signal. Moreover, we have adopted the four-stage framework of biomedical signal processing to analyze the acquired PPG signal. Before using the Savitzky-Golay derivation to pre-filter the signal and prepare it for feature extraction, it was normalized using Standard Normal Variate (SNV). In addition, four different machine learning models, including Support Vector Machine (SVM), Weighted K-Nearest Neighbors (KNN), Wide Neural Network, and Decision Tree were used for blood glucose level classification. For this study, a dataset was created consisting of 106 data, gathered from 27 subjects. Findings revealed that the Weighted KNN exhibited the best performance among other classification models, having 90.5% accuracy.

Keywords