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

Glycated hemoglobin (HbA1c) is a central biomarker for long-term glycemic control and diabetes management, traditionally quantified using laboratory-intensive chromatographic or immunochemical assays. As the global burden of diabetes continues to rise, there is growing interest in alternative, scalable approaches capable of rapid biochemical assessment. Fourier-transform infrared (FTIR) spectroscopy offers a reagent-free method that captures molecular signatures of protein glycation, but translating complex spectra into clinically interpretable HbA1c values requires robust analytical frameworks. Here, we present a complementary multi-model strategy for predicting HbA1c from FTIR spectra of whole blood. Using 685 blood samples with matched reference HbA1c measurements, we evaluated three analytically distinct yet synergistic approaches: partial least squares regression (PLSR), peak-resolved curve fitting based on pseudo-Voigt functions combined with H2O AutoML, and a convolutional neural network (CNN). PLSR and CNN models were trained on biologically informative spectral regions (800-1800 cm-{superscript 1} and 2800-3400 cm-{superscript 1}), while curve fitting focused on the fingerprint region (1000-1720 cm-{superscript 1}) to extract interpretable biochemical parameters. PLSR achieved the highest predictive accuracy (R{superscript 2} = 0.76), closely followed by the CNN (R{superscript 2} = 0.73), reflecting their ability to capture global linear and nonlinear spectral relationships. Although curve fitting yielded lower predictive performance (R{superscript 2} = 0.59), its peak-level decomposition enabled mechanistic interpretation of glycation-related changes. Explainable AI analysis using SHAP identified lipid- and protein-associated vibrations, carbohydrate-linked glycation bands, and amide-region structural features as key contributors to HbA1c prediction. Rather than treating these approaches as competing alternatives, our results demonstrate that their integration provides a more informative framework than any single model alone. By combining predictive performance with biochemical interpretability, this multi-model FTIR strategy highlights a scalable and mechanistically grounded pathway toward non-invasive HbA1c assessment and broader metabolic screening in diabetes monitoring. The code for this study is freely available at https://github.com/MelnychenkoM/ftir-hba1c-prediction.