A Multi-Wavelength NIR Spectroscopy Approach for Non-Invasive Blood Glucose Monitoring Using Artificial Neural Networks

Abstract

Diabetes mellitus (DM) is a metabolic disorder characterized by elevated blood glucose levels. Diabetic patients must frequently monitor their blood glucose (BG) levels in order to estimate the insulin intake. Invasive methods are commonly used for blood glucose monitoring because they are extremely precise but require a blood sample, which is painful and increases the risk of infections. As an alternative, noninvasive techniques don't involve any damage to the skin, making them simple, painless, and practical for regular monitoring. A non-invasive glucose monitoring system based on near-infrared (NIR) spectroscopy is proposed. The system operates at three wavelengths (940 nm, 1050 nm, and 1300 nm) and utilizes both reflectance and transmittance signals for improved accuracy. An optoelectronic setup comprising light emitting diode (LED) and photodetectors is used to acquire NIR signals, which are then processed using an ATmega32 microcontroller. Data were collected from 62 human participants (37 male, 25 female) under fasting and postprandial conditions.  An optimized artificial neural network (ANN) with 10 hidden layers achieved superior performance with an R² of 0.97 and MSE of 137.95, demonstrating strong potential for practical deployment. Clarke Error Grid analysis confirms that the 
majority of predictions fall within Zone A, indicating high clinical accuracy. The results validate the effectiveness of the proposed ANN-based approach for accurate, non-invasive, and real-time glucose monitoring.