Interpretable Random Forest (machine learning) analysis of two-trace 2D-COS FTIR spectra for multilayer biochemical characterization of colorectal precancerous.

This study investigates colorectal precancer-associated biochemical alterations across the mucosa, submucosa, and muscular layers by correlating two-trace two-dimensional Fourier transform infrared (2T2D-FTIR) spectra using random forests machine learning. The 2T2D-FTIR analyses markedly enhanced spectral resolution, enabling clear discrimination of overlapping vibrational bands that are indistinct in conventional one-dimensional FTIR spectra. Premalignant tissues exhibited pronounced reductions in lipid-associated CH2 and CH3 stretching vibrations, indicating disrupted lipid metabolism, increased lipid consumption, and membrane remodelling. Precancer-associated modifications in the amide I region indicate shifts in protein secondary structure. Substantial changes in the fingerprint-region bands reflect metabolic reprogramming, abnormal cell proliferation, and increased DNA abundance. Random Forest classification demonstrated exceptional performance, achieving perfect accuracy, precision, recall, and specificity across tissue layers. Submucosa samples showed complete separability driven by a single dominant spectral feature, while muscular-layer samples required minimal additional complexity due to slight biological heterogeneity. The generated decision trees demonstrated complete leaf purity, indicating the absence of classification ambiguity. These results underscore the strong biological consistency of premalignant spectral signatures. Overall, integrating 2T2D-FTIR spectroscopy with machine learning provides a powerful and reliable framework for probing molecular changes in colorectal cancer and supports its potential for highly accurate, non-invasive diagnostic applications.