Decoding Ascitic Immunological Niches with Multi-modal Machine Learning Reveals Prognostic and Chemoresistant Determinants in Ovarian Cancer

Lin Yang¹Tianhui He¹Jing Wang²Xiaolan Zhang¹Lin Zeng¹Qinkun Sun¹Yuelin Song¹Yufei Nie¹Xinran Gao¹Chunliang Shang^1*Hongyan Guo^1*

• ¹Peking University Third Hospital, Beijing, China
• ²Mass General Brigham, Somerville, United States

Background: Malignant ascites in high-grade serous ovarian cancer (HGSOC) represent a fluid extension of the tumor microenvironment, embedding immune programs that may inform prognosis and treatment response. We investigated whether ascitic T-cell phenotypes, integrated with clinical variables, improve prediction of overall survival (OS), progression-free survival (PFS), progression-free interval (PFI), and platinum-based chemotherapy resistance (P-DCR). Methods: We retrospectively analyzed 87 FIGO III/IV HGSOC patients with treatment-naïve ascites treated at Peking University Third Hospital (May.2019-Mar.2024; median follow-up 33 months). Ascites (>1000 mL) underwent standardized processing and multiparametric flow cytometry to quantify T-cell subsets. To prevent information leakage, we used repeated nested cross-validation with event-stratified folds: inner folds performed endpoint-specific screening with Benjamini–Hochberg FDR control, redundancy reduction, and multicollinearity checks; clinical covariates were added by incremental contribution testing. Cox proportional hazards , Random Survival Forests (RSF), and DeepSurv modeled survival endpoints; a random-forest classifier modeled P-DCR. Performance was summarized on outer folds (C-index for survival; ROC-AUC for P-DCR). Model interpretability used SHAP. Results: Across endpoints, combined clinical + ascites features outperformed single-source features, with RSF consistently best. Outer-fold testing C-indices for RSF with combined features were 0.72 (OS), 0.70 (PFS), and 0.74 (PFI). The P-DCR classifier achieved a mean AUC of 0.69 with combined features (accuracy 0.66; sensitivity 0.70; specificity 0.62). Feature-count sensitivity analyses showed performance gains plateauing at modest k (≈5–7). Kaplan–Meier curves derived from combined-feature risk scores demonstrated clear stratification. SHAP analyses indicated protective effects of PARP inhibitor maintenance across endpoints, while ascitic T-cell subsets-including PD-1⁺CD57⁺CD4⁺ and CCR7⁻CD45RA⁺CD4⁺ populations-were repeatedly associated with higher risk; age contributed strongly to PFI. Conclusions: Integrating ascitic immunophenotyping with clinical factors improves risk prediction in HGSOC, with RSF offering robust performance under rigorous, leakage-safe validation. Ascites-resident T-cell states provide complementary, reproducible prognostic signals for survival and platinum response, supporting their potential utility for patient stratification and hypothesis generation for immunomodulatory strategies.