Dual-mode NIR spectroscopy integrating characteristic wavelengths and broadband spectra for non-invasive glucose measurement

Abstract

Non-invasive blood glucose measurement technology, with its advantages of convenience and painlessness, shows promise for replacing traditional invasive measurement methods. However, existing techniques still face challenges, including limited accuracy and environmental adaptability. We present a dual-mode near-infrared (NIR) spectroscopy system for blood glucose measurement, featuring: a broadband spectrum measurement mode (900–1,700 nm/1,350–2,150 nm) and a characteristic wavelength measurement mode (940 nm, 1,050 nm, 1,310 nm, and 1,550 nm). Using this system, we collect 1,545 NIR spectral datasets for blood glucose analysis. Different prediction models are compared on the dataset, and the optimal model is selected and deployed in the system. Results demonstrate: In broadband spectrum mode, the Support Vector Regression (SVR) model achieves optimal predictive performance on both the 900–1,700 nm and 1,350–2,150 nm datasets, achieving a Mean Absolute Relative Differences (MARD) of 14.5% and 9.7% respectively. In characteristic wavelength mode, while Random Forest Regression (RFR) shows the best predictive performance with an MARD of 11.0%, the Polynomial Regression (PR) model is ultimately selected for system deployment due to practical implementation considerations, achieving an MARD of 12.8%. In all prediction results, more than 95% of the data points fall within the clinically acceptable error range (zone A and B) of the Clarke Error Grid (CEG), demonstrating strong measurement performance.