BHA: a baseline harmonization algorithm for robust classification in Raman spectroscopy

We present a new baseline correction algorithm for Raman spectroscopy aimed at improving classification robustness across a range of biomedical and analytical applications. Baseline artifacts—such as those caused by fluorescence or scattering—can obscure true Raman peaks and reduce model performance. Our method aligns inter-sample baselines using a novel combination of smoothing splines and polynomial modeling, while preserving class-relevant spectral features. We benchmarked its performance using synthetic datasets that simulate realistic variability in baseline shape, noise, and intra-group heterogeneity. Compared to established methods such as SNIP and EMSC, our approach achieves consistently higher classification accuracy and reduced residual baseline variation, especially under challenging signal-to-background conditions. These results suggest its utility as a reliable preprocessing step in Raman-based diagnostics.