Natalia Starodubtseva

Machine Learning Framework for Ovarian Cancer Diagnostics Using Plasma Lipidomics and Metabolomics

Ovarian cancer (OC), the third most common gynecologic malignancy, exhibits distinct metabolic alterations that could enable early detection via liquid biopsy. We developed an advanced machine learning pipeline integrating lipidomics (HPLC-MS, positive/negative ion modes) and NMR-based metabolomics to analyze plasma samples from 229 subjects, including 103 serous OC patients, 107 benign cases, and 19 healthy controls. By systematically evaluating feature selection methods and machine learning architectures, we identified optimal biomarker combinations for OC detection. Convolutional Neural Network (CNN) model based on Mann–Whitney-selected features demonstrated strong discriminatory power (81% accuracy) in distinguishing malignant from benign cases, while Extreme Gradient Boosting (XGBoost) combined with Support Vector Machine-Recursive Feature Elimination (SVM-RFE) achieved exceptional performance (96% accuracy) in differentiating benign from control samples. For multiclass classification, XGBoost with Kruskal–Wallis-selected features achieved 77% accuracy, while one-versus-one CNN models utilizing Mann–Whitney-selected features attained 78% accuracy, demonstrating optimal performance among tested approaches. The complementary strengths of deep learning and ensemble methods underscore their potential for tailored diagnostic applications. While clinical implementation requires further standardization, these findings provide both a methodological framework for metabolic biomarker discovery and biological insights into OC pathophysiology, paving the way for integrated multi-omics approaches in gynecologic oncology.

Raman Spectroscopy of Cell-Free Cervicovaginal Lavage for HPV Lesion Diagnosis: A Pilot Study

High-risk human papillomavirus (HPV) is the leading etiological factor in cervical cancer, creating a pressing need for less invasive and more objective diagnostic tools. This pilot study pioneers the application of Raman spectroscopy to cell-free cervicovaginal lavage (CVL) for distinguishing between low-grade and high-grade squamous intraepithelial lesions (LSIL and HSIL) in HPV-positive patients. Raman spectra were acquired at 532-nm excitation from cell-free CVL samples of 20 patients with histologically confirmed LSIL (n = 9) or HSIL (n = 11). Comparative analysis of Raman bands revealed a significant biochemical shift in HSIL, presumably characterized by reduced glycogen and lactate/lactic acid levels alongside substantially elevated heme proteins. A diagnostic model based on key spectral intensity ratios achieved differentiation between LSIL and HSIL with 80% sensitivity and 86% specificity. These findings demonstrate that Raman spectroscopy of cell-free CVL effectively captures profound metabolic and microvascular alterations characteristic of neoplastic progression, showcasing its strong potential as a rapid, cost-effective, non-invasive, and objective tool for cervical lesion risk stratification.