Blood-Glucose Testing: AI and FT-IR Claim Improved Accuracy to 98.8%

A study led by Liying Song, Zhiqiang Han, Po-Wan Shum, and Woon-Ming Lau has successfully transformed the accuracy of non-invasive blood-glucose testing. Conducted at Linyi University and the University of Science and Technology Beijing, the research integrates multiple technological advancements to overcome limitations in traditional Fourier transform-infrared (FT-IR)-based glucose testing (2).

Blood-Glucose Testing

Current non-invasive glucose testing methods rely on attenuated total reflection (ATR) combined with infrared spectroscopy, specifically ATR FT-IR, where infrared light is reflected from sample tissue using the ATR phenomenon to detect low signal glucose levels in a complex background. However, this approach struggles with sensitivity and interference from other biomolecules and from optical differences in tissues. To address these challenges, the research team implemented three key innovations: a multi-pass ATR reflections setup, quantum cascade laser (QCL) illumination, and advanced machine learning (ML) data analysis techniques (2).

Upgrading FT-IR for Superior Sensitivity

The first enhancement involved replacing the conventional single-pass ATR setup with a multiple-reflection ATR (MATR) configuration. This modification increased the signal-to-noise ratio by a factor of two to three, allowing for clearer glucose detection in complex biological tissue samples (2).

The second major enhancement replaced the traditional infrared lamp with a QCL, further tripling the signal-to-noise performance. The QCL, operating between 935 and 1128 cm⁻¹, precisely targeted glucose’s mid-infrared spectral signature, improving accuracy and reliability.

Leveraging Two-Dimensional Correlation Spectroscopy and AI

Beyond hardware improvements, the study leveraged two-dimensional correlation spectroscopy (2D-COS) to refine spectral analysis. Unlike traditional one-dimensional spectral methods, 2D-COS analyzes spectral variations across two wavelength axes, effectively minimizing interference from overlapping biomolecular signals (2).

To maximize analytical precision, the researchers also incorporated ML-based decision trees for hyperglycemia screening. They tested multiple algorithms, with a final approach combining three decision-tree models and a majority-voting mechanism. This strategy significantly improved classification performance, achieving an unprecedented 98.8% accuracy in distinguishing diabetic from non-diabetic glucose levels (2).

Extensive Testing and Validation

The study’s rigorous validation process involved 24 glucose concentrations (70–300 mg/dL) spanning both non-diabetic and diabetic ranges. Each concentration underwent 100 repeated spectral measurements, totaling 7,200 test spectra. The MATR-QCL-2D-COS combination yielded the best results, demonstrating not only high accuracy but also consistency across samples (2).

Implications for Diabetes Management

This research marks a notable step toward practical, non-invasive glucose monitoring solutions. By surpassing the accuracy benchmarks set by previous studies, including those requiring mechanical pressure for sampling, the upgraded FT-IR method could soon become a viable alternative to traditional blood-based glucose tests (2).

The integration of ML with advanced spectroscopic techniques opens new possibilities for rapid, painless, and highly accurate diabetes screening. As non-invasive technologies continue to advance, this breakthrough may pave the way for real-world applications in continuous glucose monitoring and early diabetes detection.

References

(1) National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) Diabetes Home Page. https://www.niddk.nih.gov/health-information/diabetes (accessed 2025-01-28)

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