Non-Invasive Continuous Real-Time Blood Glucose Estimation Using PPG Features-based Convolutional Autoencoder with TinyML Implementation

In this paper, we developed a convolutional autoencoder for non-invasive continuous monitoring of blood glucose levels (BGL) using four photoplethysmography (PPG) features. The model was specifically designed to account for temporal relations among consecutive PPG segments' features and transient outliers encountered in real-time operation. By means of Tiny Machine Learning (TinyML), the model was embedded in an edge device, Arduino Nano 33 BLE Sense, for real-time continuous predictions of BGL. On a PC, the model was tested using a public dataset of 33 subjects and achieved a mean absolute error (MAE) 5.55 mg/dL, standard error of prediction (SEP) 7.18 mg/dL, and 97.57% success rate in zone A of Clarke error grid (CEG). On the edge, the model was tested on new 8 subjects and obtained a MAE 5.16 mg/dL and 100% of predicted BGL falling into zone A. Overall, the integration of the proposed model and the feature set resulted in substantial gains in terms of applicability, effectiveness, efficiency, and interpretability on both cloud and edge infrastructures.