FTIR-based profiling of sinusitis: Stabilized feature selection through Boruta and Monte Carlo methods

Sinusitis is a prevalent inflammatory condition lacking rapid, non-invasive diagnostic tools. In this study, Fourier-transform infrared (FTIR) spectroscopy was applied to distinguish serum samples collected from sinusitis patients and healthy controls. FTIR spectra were preprocessed and analyzed using a robust, two-stage feature selection framework combining the Boruta algorithm and Monte Carlo Feature Selection within a nested cross-validation design. Both methods consistently highlighted a narrow spectral band between 3400 and 3550 cm^-1 as a key region for discrimination. However, derivative spectra outside this band (e.g. 1940-2360 cm^-1) were also repeatedly selected, indicating dispersed biochemical signal. Eighteen consensus features (first-derivative, full-range) enabled an SVM to reach 94 % accuracy (sens 1.00, spec 0.83, κ 0.87) on the 17-patient held-out set. Although overlap among individual 5 cm^-1 bins is only moderate (mean Jaccard ≈ 0.26), a bin from the 3400-3550 cm^-1 band emerges in virtually every cross-validation split, implying that this high-wavenumber region robustly carries disease-related biochemical signal. Feature selection frequency heatmaps emphasized this region's importance across cross-validation folds. Furthermore, dimensionality reduction techniques (PCA, UMAP, and t-SNE) revealed distinct clustering between patient and control groups, and permutation testing confirmed the statistical significance of the classification. These findings demonstrate that serum-based FTIR spectroscopy, combined with rigorous feature selection, holds promise as a non-invasive diagnostic aid for sinusitis, while highlighting the need for broader comparative studies in future research.