Bayesian Optimization-Based CNN Model for Blood Glucose Estimation Using Photoplethysmography Signals

This paper presents a novel Bayesian optimization-based convolutional neural network (CNN) model for non-invasive blood glucose level (BGL) predication using photoplethysmography (PPG) signals. The Bayesian search found the optimal CNN architecture, achieving 92.85% accuracy in Clarke error grid (CEG) zone A with a wide detection range of 50–200 mg/dL and without any projected values within zones C, D, or E. This indicates that the suggested model is clinically acceptable. The model demonstrated better results in performance assessment using mean absolute error (MAE) and root mean squared error (RMSE) measures. This was achieved by utilizing only the features extracted by the CNN model, eliminating the need for additional feature extraction. This approach reduces computational demands and improves real-time performance. Also, the model achieved a much lower Standard Error of Prediction (SEP) compared with ML-based models. The proposed model is lightweight and can be easily deployed as a stand-alone device using a microcontroller like the Arduino Nano 33 BLE Sense connected with a PPG sensor.