Photoplethysmography Based Blood Glucose Estimation Using Convolutional Neural Networks

This paper introduces the design and measurements for non-invasive blood glucose level (BGL) estimation using a convolutional neural network (CNN) based on photoplethysmography (PPG). The prototype consists of a PPG sensor connected to a microcontroller (MCU) Arduino Nano 33 BLE Sense. The PPG-only based CNN model deployed on the MCU showed 89.28% of the predicted samples in zone A of a Clarke error grid (CEG). When the mean power spectrum feature from PPG signals was included, the results demonstrated an improvement in the accuracy to be 92.85%. The proposed system is real-time and non-invasive that can be used to replace the existing invasive glucometers.