FTIR spectroscopy combined with machine learning reveals molecular signatures distinguishing three phenotypes of endometriosis.

Endometriosis is a chronic inflammatory disorder in which endometrial tissue grows outside the uterus, leading to pelvic pain and infertility. It remains a major challenge in women's health due to delayed diagnosis and limited therapeutic outcomes. The disease is heterogeneous and manifests in three main phenotypes: superficial peritoneal lesions, ovarian endometriomas, and deep infiltrating endometriosis, the latter still poorly understood at the molecular level. In this study, Fourier-transform infrared (FTIR) spectroscopy combined with machine learning was applied to explore biochemical differences among these phenotypes. Spectral analysis demonstrated that lipid- and carbohydrate-associated vibrations, progressively increased from ovarian to bowel and were most pronounced in peritoneal endometriosis. These findings suggest phenotype-specific alterations in lipid and carbohydrate composition. Using the Boruta feature selection algorithm, discriminative spectral intervals were identified for pairwise classification. Key signals involved protein, lipid, and carbohydrate vibrations, including CO and C-O stretching and N-H bending. Machine learning models: Deep Learning (DL), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost), were trained on both full spectra and Boruta-reduced datasets. Extending to three-class classification, Boruta-selected features enabled robust discrimination across all phenotypes. A decision tree revealed the 1752 cm-1 ester CO band as a critical marker for phenotype differentiation. Overall, FTIR spectroscopy combined with machine learning provides valuable molecular insights into endometriosis and represents a powerful tool for distinguishing its clinical phenotypes.