Precision Signal Boost for Non-Invasive Blood-Glucose Tests with Advanced FT-IR and Machine Learning

Introduction

Diabetes monitoring remains heavily reliant on invasive finger-prick blood tests, despite decades of research, publications, patents, and product prototypes into non-invasive alternatives. A recent study published in Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy reports significant accuracy improvements in non-invasive glucose detection using an upgraded approach to Fourier transform infrared spectroscopy (FT-IR) (1).

The work, authored by L. Song, Z. Han, P.W. Shum, and W.M. Lau, from Linyi University, China and University of Science and Technology in Beijing, China, demonstrates that carefully designed enhancements to FT-IR—paired with advanced data-mining and machine learning—can elevate non-invasive glucose tests to levels of accuracy often previously thought to be unattainable (1).

Blood-glucose variation testing attempted using advanced FT-IR and machine learning © petrroudny -chronicles-stock.adobe.com

Upgrading FT-IR for Higher Sensitivity

Traditional FT-IR systems often rely on a single-pass attenuated total reflection (ATR) crystal, which limits sensitivity when detecting glucose in biological tissue or fluids. According to the study, this limitation was addressed by replacing the ATR with a multi-pass multiple attenuated total reflection (MATR) setup (1).

The MATR prism, made of ZnSe and designed with seven reflection passes, increased the signal-to-noise ratio by two to three times without altering the spectral fingerprints of glucose. This improvement provided stronger absorption features necessary for more reliable analysis (1).

Introducing the Quantum Cascade Laser

The research team further improved performance by replacing the conventional infrared lamp in the Nicolet iS50 FT-IR spectrometer with a quantum cascade laser (QCL). The QCL boosted the signal-to-noise ratio by an additional factor of three, producing clearer spectra of glucose solutions mixed with bovine serum albumin (1).

This enhancement proved critical for distinguishing glucose levels across a clinically relevant concentration range of 70–300 mg/dL, which spans both non-diabetic and diabetic conditions (1).

Two-Dimensional Correlation Spectroscopy

While stronger signals improve clarity, overlapping spectral features from biological molecules still complicate analysis. To address this, the team employed two-dimensional correlation spectroscopy (2DCS). Unlike conventional one-dimensional spectral analysis, 2DCS compares spectral intensity correlations across two wavelength axes (1).

The study found that 2DCS suppressed interference from non-glucose molecules and, when paired with the MATR-QCL setup, raised classification accuracy for hyperglycemia detection to 96.3% (1).

Machine Learning for Decision Accuracy

The final stage of the study applied advanced machine learning (ML) to maximize decision reliability. Three decision-tree methods were used to generate trial outcomes, followed by a “majority-voting” mechanism that produced a final classification (1).

Using random forest (RF) ML algorithms, the system achieved a reported accuracy of 98.8% in distinguishing hyperglycemia from normoglycemia. The results were validated through 100 spectral measurements for each of 24 glucose concentrations, representing both diabetic and non-diabetic ranges (1).

Clinical Relevance

Earlier work has demonstrated that careful sampling pressure on the ATR crystal could meet U.S. Food and Drug Administration (FDA) accuracy benchmarks for non-invasive glucose testing. However, the new results indicate that with the MATR-QCL-2DCS approach, high accuracy can be achieved without the need for engineered sampling pressure (1,2).

This combination of spectroscopic enhancements and ML represents one of the most precise demonstrations of non-invasive glucose detection to date, with potential to bring the technology closer to clinical adoption (1).

Conclusion

By integrating multiple reflections, a quantum cascade laser source, two-dimensional correlation spectroscopy, and machine learning, this study demonstrated that non-invasive FT-IR glucose monitoring can achieve near-clinical accuracy. With a reported success rate of 98.8%, the work points to new possibilities for diabetes management without the need for blood extraction (1).

References

(1) Song, L.; Han, Z.; Shum, P. W.; Lau, W. M. Enhancing the Accuracy of Blood-Glucose Tests by Upgrading FT-IR with Multiple-Reflections, Quantum Cascade Laser, Two-Dimensional Correlation Spectroscopy and Machine Learning. Spectrochim. Acta, Part A 2025, 327, 125400. DOI: 10.1016/j.saa.2024.125400

(2) Chen, J. Y.; Zhou, Q.; Xu, G.; Wang, R. T.; Tai, E. G.; Xie, L.; Zhang, Q.; Guan, Y.; Huang, X. Non-Invasive Blood Glucose Measurement of 95% Certainty by Pressure Regulated Mid-IR. Talanta 2019, 197, 211–217. DOI: 10.1016/j.talanta.2019.01.034