Rapid serum-based differentiation of ovarian tumors and assessment of post-treatment disease-free status using portable Raman spectroscopy and machine learning.

Rapid, low-cost tools for ovarian tumor discrimination and post-treatment assessment remain limited because conventional serum biomarkers such as CA125 have suboptimal sensitivity and specificity. Serum contains proteins, lipids, carbohydrates, and nucleic acids that generate disease-related spectrochemical fingerprints, which portable Raman spectroscopy can capture with minimal sample preparation. We investigated serum Raman profiling for ovarian tumor discrimination and cross-sectional characterization of post-treatment disease-free status. Serum samples from 280 participants (85 benign, 82 malignant, and 113 independent follow-up patients) were analyzed over 400-1800 cm⁻1. Candidate regions were re-selected in each training set and represented by local mean intensity or an 11-point local spectral profile (Raw11). Four classifiers were evaluated using fivefold cross-validation, repeated nested fivefold cross-validation, 100 stratified 70:30 holdout repetitions, and label-permutation testing. Raw11 consistently outperformed mean intensity. RBF-SVM showed balanced discrimination and stability (AUC = 0.969, accuracy = 0.940, sensitivity = 0.941, specificity = 0.939), with a repeated holdout AUC of 0.943 ± 0.054 and permutation P = 0.001996. Representative Raman regions were not significantly correlated with continuous CA125, indicating that Raman features were not simple surrogates for CA125. Follow-up samples were strongly distinguishable from active malignancy (RBF-SVM AUC = 0.996) but less separable from benign samples (random forest AUC = 0.840), and frozen benign-malignant models placed them closer to benign spectral space. These findings support portable serum Raman spectroscopy as a rapid, label-free adjunct for ovarian tumor discrimination and cross-sectional assessment of post-treatment disease-free serum phenotypes.