A Multidimensional Data-Driven Approach to Surgical Plan Optimization and Postoperative Residual Tumor Prediction in Ovarian Cancer

Lin Yang¹Tianhui He¹Ping Yu¹Chunliang Shang¹Jing Wang^2*Qinkun Sun¹Yi Xie¹Jianling Yang^1*Hongyan Guo^1*

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

Backgrounds: Ovarian cancer represents a deadly gynecological malignancy, with surgical treatment being a key component of its management. We sought to integrate clinical characteristics and ascites immune microenvironment features into a deep learning model to predict postoperative residual tumor status and assist in surgical decision-making. Methods: 118 FIGO III/IV high-grade serous ovarian cancer (HGSOC) patients treated at Peking University Third Hospital (2019-2024) were enrolled. Clinical characteristics, surgical methods, and postoperative residual tumor status were collected. Ascites samples were processed via density gradient centrifugation and flow cytometry. Deep learning model was built by fusing clinical and immune data, and its performance was validated under a gradient of feature quantities (5-45 features) to optimize feature selection. Model performance was comprehensively evaluated on a test set (20% of the dataset) using metrics including accuracy, precision, recall, and F1 score, and compared with traditional machine learning models (random forest, XGBoost, et al). Confusion matrices and probability heatmaps were used for visual analysis.For model interpretability, we presented feature importance and results from SHAP analysis. Results: Our model achieved 70.83% accuracy, 71.21% precision, 70.83% recall, and 70.89% F1 score on the test set, outperforming traditional machine learning models: random forest (accuracy: 64.6%, precision: 65.1%, recall: 64.6%, F1 score: 66.4%), XGBoost (accuracy: 66.7%, precision: 67.0%, recall: 66.7%, F1 score: 66.6%), and logistic regression (accuracy: 58.3%, precision: 59.0%, recall: 58.4%, F1 score: 58.2%). It demonstrated strong performance in identifying high-risk R2 cases but showed limitations in differentiating between R0 and R1 statuses. Probability heatmaps visualized the distribution of R0, R1, and R2 probabilities under different surgical methods, facilitating intuitive clinical reference. Interpretability analysis via permutation feature importance and SHAP highlighted the critical role of surgical methods and specific immune microenvironment features in predictive outcomes. Conclusion: This study developed a novel deep learning-based model to predict postoperative residual tumor probability, integrating clinical and immune microenvironment data. While the model excelled in identifying high-risk cases (e.g., R2), further optimization is needed to improve R0 and R1 differentiation. Future research should expand datasets and integrate multi-omics data to enhance predictive accuracy and clinical applicability.