Edge Classification of Non-Invasive Blood Glucose Levels Based on Photoplethysmography Signals

Abstract

Diabetes monitoring systems are critical for avoiding potentially significant medical bills. Only invasive methods are currently on the market. These procedures have substantial drawbacks since they cause patients uncomfortable while collecting blood specimens. An approach checking blood glucose levels (BGL) that is comfortable, continuous, and non-injury will become a new alternative to invasive procedures. Photoplethysmography can identify cardiovascular disease. Because of these qualities, PPG signals directly impact diabetes patients. Edge Computing (EC) is a relative newcomer to handling modern challenges more efficiently through machine learning. In this study, Edge Impulse uses the TensorFlow environment to train, optimize, and deploy machine learning models to embedded devices. The study examines three different forms of raw data used as inputs. We look at the original PPG signal, the PPG signal with instantaneous frequency, and the PPG signal with spectral entropy. The data set was created using Guilin People's Hospital's public database, including 219 people. The ages represented in the data set range from 20 to 89 years. According to the findings of the model testing, the PPG signal with instantaneous frequency shows the best results.