Development of a Non-Invasive Blood Glucose Monitoring Device Using Machine Learning Technology

Blood glucose monitors are critical to diabetes management. There is no permanent medicine to cure diabetes. Presently, invasive glucose blood meters extract blood sample by pecking a needle into the patient’s fingers. This results in the formation of copious calluses on the fingertips and causes more pain to lure blood again and again for repetitive measurements.
The study aims to develop Non-Invasive Glucometer to monitor the glucose level of a person using a Wi-Fi module. A variation in amplitudes, and phases of received packets, helps to measure glucose levels. A Hampel filter is used to suppress abrupt amplitude variations occurring due to environmental effects. Further, the Fast-Tree Regression algorithm is used to train the model for different glucose concentrations for accurate prediction and detection of diabetes. It also reduces dataset dimension for minimizing the training time of the device. Thereafter, Clarke Error Grid Analysis helps to estimate the accuracy.
Two ESP32 Wi-Fi devices, are installed on a computer for real time sensing of Channel State Information (CSI) between an Receiver Access Point and Transmitter Station. Further, additional header information such as MAC address, RSSI, and other metadata along with the CSI is sent for all 64 subcarriers. Here, statistical regression analysis is considered only to confirm the results.
The accuracy achieved is 95 % with coefficient of determination in terms of an R2 value of 0.99. The device measures glucose level in less than 3 sec. It can store 2000 test results with time, date, daily and weekly average reports for random, before, and after the meal. A containers containing air and 5% Glucose solution helps to validate the models behavior with specific glucose content.
The portable painless device is found to be useful to monitor the glucose level at home and office. The benefits is low-cost and Non –Invasive.