Non-Invasive Glucose Monitoring using Machine Learning

Abstract

Diabetes Mellitus (DM), the metabolic disorder can cause serious health issues if not managed properly. The conventional blood glucose monitors are invasive and causes pain and discomfort to patients. Therefore, the study was aimed to measure blood glucose non-invasively based on a machine learning technique. This system involves a Photoplethysmograph (PPG) based system using a light source of wavelength 525nm, 660nm, and 950nm to determine the blood glucose parameters. The light source illuminates, the skin at the wrist and the reflected beam is captured by a photodiode receiver. The detected signal is conditioned, digitalized and passed to the Arduino UNO microcontroller. The Arduino board derives the PPG signal in accordance with the subject blood glucose values. The raw waveform is pre-processed and subsequently segmented for obtaining the peak of the PPG signal. The random forest machine learning technique is implemented on the acquired segmented signal, to obtain various statistical features namely mean, variance, skewness, entropy, kurtosis and standard deviation. The machine learning system is designed and trained to estimate blood glucose from the extracted features. The blood glucose obtained from the proposed method, were comparable with the standard method. The results studied infers that the proposed method can be used for blood glucose monitoring.