Spectroscopic Quantification of Plasma Free Hemoglobin Based on Paired Domain Adaptation and Orthogonality Constraints

Abstract

Objective

We propose a paired domain-adaptation deep regression method for multi-pathlength spectroscopy to quantify plasma free hemoglobin (FHB) robustly across measurement conditions.

Methods

UV–Vis–NIR spectra (300–1160 nm; 945 wavelengths) were acquired using an Avantes spectrometer, with five optical pathlengths per sample. Spectra were preprocessed by standard normal variate (SNV), and labels were log-transformed (log (1 + y)) to mitigate long-tailed instability. The network integrates domain–path affine calibration, a 1D-CNN encoder, and attention-based multi-path fusion, followed by shared–private feature disentanglement. A paired consistency loss aligns only the shared representation across paired domains, and an orthogonality constraint encourages domain-specific separation. Performance was evaluated via regression-stratified five-fold cross-validation using RMSE and R² on the raw scale.

Results

For N = 251 samples, λ_pair = 1.0 achieved RMSE = 260.53 ± 62.01 and R² = 0.748 ± 0.088.

Conclusion

The method improves cross-domain robustness and interpretability for plasma FHB prediction.