Abstract Endometriosis is a chronic, estrogen-dependent disorder characterized by the presence of endometrial-like tissue outside the uterus, often leading to pelvic pain and infertility. Despite its high prevalence, current diagnostic methods rely mainly on invasive laparoscopy, and there is a need for novel, noninvasive approaches that can support disease detection and staging. In this study, Fourier Transform Infrared (FTIR) spectroscopy of serum was combined with machine learning (ML) techniques to evaluate whether biochemical alterations reflected in infrared spectra can differentiate between different stages of endometriosis. Serum samples obtained from patients diagnosed with stage I, II/III, and IV endometriosis were analyzed using Principal Component Analysis (PCA) to visualize spectral variance and assess clustering of samples, while supervised classification was performed using Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), and Tabular Prior-Data Fitted Networks (TabPFN) models. Model validation was conducted using a repeated, stratified cross-validation approach (10 repetitions, 10 splits). The RF algorithm was applied for feature selection to identify the most informative wavenumbers, followed by decision tree analysis to determine the optimal spectral discriminants. The results revealed distinct biochemical trends across disease stages: early endometriosis (stage I) exhibited higher absorbance in the amide I–II and phosphate regions, suggesting a greater contribution of protein- and phospholipid-associated spectral features, while advanced stages (IV) showed increased lipid-associated Csingle bondH stretching bands, consistent with lipid-associated biochemical remodeling and inflammatory activity. Machine Learning models achieved the highest classification performance (ROC AUC > 0.96 for comparison I vs IV stages of endometriosis), also when trained on selected subset of features, confirming the discriminative power of specific wavenumbers. It should be noted that all analyses were performed on a single patient cohort from one clinical center, and model validation was conducted using internal cross-validation only. Therefore, these results represent a proof-of-concept study and further external validation is required before clinical translation. Decision tree analysis identified key differentiating bands at: 835 cm−1 for differentiation stages I and II/III, 1209 cm−1 for distinguishing stages I and IV and at 980 cm−1 for differentiation stages II/III and IV. Overall, the integration of FTIR spectroscopy with machine learning demonstrates a strong potential for noninvasive differentiation of endometriosis stages, offering a promising biochemical fingerprint-based approach for improved diagnosis and disease monitoring.