FEEDFORWARD BACKPROPAGATION MODEL OF NEURON FOR BLOOD GLUCOSE ESTIMATION

Abstract

One possible non-invasive technique for determining glucose levels is near-infrared (NIR) spectroscopy, but the accuracy of predictive models remains a challenge due to spectral noise and nonlinearity. This study explores a Feedforward Backpropagation Neural Network (FFBPNN) trained with Bayesian Regularization (BR) to enhance prediction accuracy and generalization. The BR training function optimizes the network by minimizing overfitting, adjusting weight distributions, and improving robustness against spectral variations. Preprocessing techniques such as Savitzky-Golay filtering, moving average filtering are integrated to extract meaningful time and frequency domain features. The proposed FFBPNN-BR model achieves superior accuracy compared to conventional training methods, with improved Clarke error grid analysis results. This approach demonstrates significant potential for real-time, non-invasive glucose monitoring, advancing wearable and clinical diagnostic applications. Future work focuses on edge deployment and multimodal sensor fusion for enhanced robustness.