📄 health informatics

Decoupling Accuracy and Explainability: Machine Learning Strategies for HbA1c Prediction and Biomarker Discovery in Blood FTIR Spectroscopy

This study demonstrates that integrating partial least squares regression, convolutional neural networks, and curve-fitting approaches on FTIR blood spectra creates a synergistic framework that simultaneously achieves high accuracy in predicting HbA1c levels and provides mechanistic interpretability of glycation-related biomarkers for scalable diabetes monitoring.

Original authors: Melnychenko, M., Makhnii, T., Midlovets, K., Dmyterchuk, B., Krasnienkov, D.

Published 2026-01-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Melnychenko, M., Makhnii, T., Midlovets, K., Dmyterchuk, B., Krasnienkov, D.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Big Picture: Listening to Blood's "Song"

Imagine your blood is a busy orchestra playing a complex song. Every instrument (proteins, fats, sugars) plays a specific note. When you have diabetes, the "sugar" section of the orchestra starts playing louder and changes the tune slightly because sugar sticks to your red blood cells (a process called glycation).

Traditionally, doctors check this by taking blood to a lab and running it through a slow, expensive machine (like HPLC) to count exactly how many "sugar notes" are stuck to the cells. This paper asks: Can we just listen to the blood's "song" using a special light scanner (FTIR) and use a computer to figure out the sugar levels instantly?

The researchers tried to answer this by teaching three different types of "computers" to listen to the blood's song and predict the sugar levels (HbA1c).

The Three "Computers" (Models)

The team didn't just use one method; they tried three different ways to interpret the data, like three different detectives solving the same mystery.

1. The "Pattern Matcher" (PLSR)

How it works: This model looks at the whole song at once. It doesn't try to identify every single instrument; instead, it looks for the overall "vibe" or shape of the sound wave. It's like a music critic who says, "This song sounds like a jazz track with a lot of sugar in it," based on the general feel.
The Result: This was the best detective. It got the most accurate predictions (about 76% accuracy). It was great at seeing the big picture.

2. The "Deep Learner" (CNN)

How it works: This is a type of Artificial Intelligence that learns by itself. It's like a baby bird that hears thousands of songs and eventually figures out the rules of music without anyone teaching it the names of the notes. It looks for hidden, complex patterns that a human might miss.
The Result: This was the second-best detective (about 73% accuracy). It did almost as well as the Pattern Matcher, proving that the computer could learn the "rules" of the blood song on its own.

3. The "Microscope" (Curve Fitting)

How it works: This method is the opposite of the others. Instead of looking at the whole song, it tries to isolate every single instrument. It breaks the sound wave down into individual peaks (notes) and measures their height and width. It's like a music teacher saying, "The violin is playing at 1030 Hz, and the drum is at 1574 Hz."
The Result: This was the least accurate at predicting the final number (about 59% accuracy). However, it was the most honest. Because it broke the song down note-by-note, the researchers could actually explain why the computer made its guess. It told them exactly which chemical parts of the blood were changing.

The "Aha!" Moment: Why Use All Three?

The paper's main discovery is that these three methods shouldn't be rivals; they should be a team.

The Pattern Matcher (PLSR) and The Deep Learner (CNN) are great at giving you the right answer (the HbA1c number), but they are a bit of a "black box." You know the answer, but you don't know exactly why the computer thought that.
The Microscope (Curve Fitting) isn't as good at guessing the number, but it acts as the translator. It tells you, "The computer guessed high because the 'sugar note' at 1030 Hz got louder, and the 'protein note' at 1574 Hz got quieter."

By combining them, the researchers got the best of both worlds: a highly accurate prediction plus a clear explanation of the biology behind it.

What Did They Find in the Blood?

Using the "Microscope" and a special tool called SHAP (which acts like a highlighter to show which notes matter most), they found that the blood's song changes in specific ways when sugar sticks to blood cells:

The Sugar Notes: Certain frequencies (around 1011 and 1030) got stronger. These are the sounds of sugar molecules attaching to proteins.
The Protein Changes: The shape of the protein "instruments" (hemoglobin) changed. Some parts of the protein folded differently (like a crumpled paper airplane), which showed up as changes in the "Amide" notes (around 1574 and 1680).
The Fat Connection: They also noticed changes in the "fat" notes (lipids), suggesting that when sugar control is poor, the body's fat metabolism gets messy too.

The Bottom Line

This study proves that you can use a quick, reagent-free light scan (FTIR) to estimate diabetes levels.

Don't pick just one tool: If you only use the "Deep Learner," you get a good guess but no explanation. If you only use the "Microscope," you get a great explanation but a weaker guess.
The Winning Strategy: Use a mix. Let the powerful AI models do the heavy lifting for accuracy, and use the detailed curve-fitting to understand the biology.

The authors conclude that this "multi-model" approach is a promising step toward a future where doctors could scan a drop of blood with a simple device and instantly get a diabetes reading that is both accurate and scientifically explainable.

(Note: The paper explicitly states this is a preprint and has not yet been peer-reviewed, so these results are currently a proof-of-concept rather than a ready-to-use medical tool.)

Technical Summary: Decoupling Accuracy and Explainability in HbA1c Prediction via FTIR Spectroscopy

Problem Statement
Glycated hemoglobin (HbA1c) is the standard biomarker for long-term glycemic control and diabetes management. Current clinical quantification relies on laboratory-intensive methods such as high-performance liquid chromatography (HPLC) and immunoassays, which require specialized reagents, instrumentation, and time, limiting scalability in low-resource settings. Fourier-transform infrared (FTIR) spectroscopy offers a reagent-free, rapid alternative that captures molecular signatures of protein glycation. However, translating complex, overlapping FTIR spectra into clinically interpretable HbA1c values remains challenging. Existing approaches often face a trade-off: statistical models like Partial Least Squares Regression (PLSR) offer predictive power but lack mechanistic interpretability, while peak-resolved methods offer biochemical insight but may suffer from lower predictive accuracy. There is a need for a framework that balances high predictive accuracy with biochemical explainability.

Methodology
The study utilized a dataset of 673 whole blood samples (derived from an initial 685 after outlier removal) collected from healthy, prediabetic, and diabetic participants. Reference HbA1c values were obtained via high-performance ionic liquid chromatography. FTIR spectra were acquired using an ATR-FTIR spectrometer (Nicolet iS50) in the range of 4000–400 cm⁻¹, with specific focus on biologically relevant regions.

The authors evaluated three distinct, synergistic machine learning strategies:

Partial Least Squares Regression (PLSR): Trained on spectral regions 800–1800 cm⁻¹ and 2800–3400 cm⁻¹. Data was preprocessed using Savitzky-Golay filtering (first derivative) and vector normalization. The model was optimized via 10-fold cross-validation to determine the optimal number of latent components (nine were selected).
Convolutional Neural Network (CNN): A 1D-CNN architecture adapted from McHardy et al., trained on the same spectral regions as PLSR. The model utilized two sequential convolutional units (64 filters, kernel size 5), followed by ReLU activation, batch normalization, max-pooling, and fully connected dense layers. Training employed the Adam optimizer with Mean Squared Error (MSE) loss.
Curve Fitting with H2O AutoML: Focused on the "fingerprint" region (1000–1720 cm⁻¹). Spectra were smoothed, baseline-corrected (rubberband), and normalized. Peak deconvolution was performed using pseudo-Voigt functions via the JAXFit package. The resulting peak parameters (position, area, height, FWHM) served as features for H2O AutoML models.
Explainable AI (XAI): SHAP (SHapley Additive exPlanations) values were applied to the H2O AutoML models to quantify the magnitude and direction of feature contributions, linking spectral peaks to biochemical mechanisms.

Key Results

Predictive Performance: The PLSR model achieved the highest predictive accuracy on the independent test set ( $R^2 = 0.76$ , RMSE = 1.16, MAE = 0.89). The CNN model performed comparably ( $R^2 = 0.73$ , RMSE = 1.22, MAE = 0.94), suggesting that the dominant spectral relationships for HbA1c in this dataset are largely linear, with nonlinear effects playing a secondary role. The Curve Fitting approach yielded lower predictive performance ( $R^2 = 0.59$ , RMSE = 1.51, MAE = 1.11) but provided high-quality spectral fits ( $R^2 > 0.9999$ ).
Biochemical Interpretability: SHAP analysis identified specific spectral features driving predictions. Key contributors included:
- Carbohydrate/Glycation bands: 1011 cm⁻¹ and 1030 cm⁻¹ (C–O/C–N stretching), associated with glucose-derived adducts and early glycation.
- Protein Secondary Structure: Amide III regions (1302–1313 cm⁻¹) and Amide I/II regions (1574, 1678, 1687 cm⁻¹), indicating structural rearrangements (e.g., $\beta$ -turns, loss of $\alpha$ -helix) in hemoglobin and serum proteins.
- Lipid/Carboxylate interactions: Peaks at 1366 cm⁻¹ and 1390–1402 cm⁻¹, linked to lipid composition and carboxylate groups of amino acids (Asp/Glu).
Model Comparison: While PLSR and CNN excelled in generalization, the Curve Fitting method successfully deconvoluted complex spectra to isolate specific glycation-sensitive peaks, validating the biochemical basis of the predictions.

Significance and Claims
The paper argues that treating these modeling approaches as competing alternatives is suboptimal. Instead, the authors propose a complementary multi-model strategy.

Integration of Strengths: The study demonstrates that integrating predictive models (PLSR/CNN) with mechanistic analysis (Curve Fitting) provides a more informative framework than any single model alone. PLSR and CNN capture the broad linear and nonlinear patterns necessary for robust clinical prediction, while peak-resolved analysis ensures biochemical interpretability and mechanistic validation.
Clinical Potential: This strategy highlights a scalable, reagent-free pathway for non-invasive HbA1c assessment. By combining predictive performance with explainability, the approach supports the use of FTIR not just as a surrogate for HbA1c, but as a versatile screening platform capable of revealing a broader biochemical profile (lipids, protein structure, nucleic acids) from a single measurement.
Limitations: The authors modestly acknowledge limitations, including the inherent constraints of FTIR (water absorption, baseline variations), the reliance on specific peak shape assumptions in curve fitting, and the need for larger, more diverse datasets to confirm generalizability across different populations.

In conclusion, the study validates that a multi-model framework can decouple the trade-off between accuracy and explainability, offering a mechanistically grounded approach to diabetes monitoring using FTIR spectroscopy.