Early Prediction and Risk Stratification of Ovarian Cancer Using Explainable AI and Machine Learning for Improved Survival Outcomes.

Primary supervisor

Jesmin Nahar

Research area

Machine Learning

1. Research Objectives & Clinical Problem Late-stage diagnosis remains the leading cause of high mortality in ovarian cancer, as early stages are often asymptomatic or present with non-specific abdominal symptoms. Standalone clinical indicators like CA125 or traditional screening algorithms (e.g., ROMA) frequently produce ambiguous results or lack sufficient sensitivity for early detection. This study aims to construct a transparent, non-invasive predictive framework combining machine learning and explainable AI (XAI) to differentiate malignant from benign pelvic masses, stratify patient risk, and advance diagnostic timelines to improve overall survival outcomes.

2. Data Preprocessing & Predictive Modeling The study utilizes multi-modal clinical and laboratory data, integrating basic demographics (e.g., age and menopausal status) with routine blood parameters, acute-phase reactants, liver function profiles, and tumor biomarkers (CA125, HE4, ROMA score). Feature selection algorithms (e.g., Boruta) and statistical imputation methods (e.g., KNN/Iterative Imputer) are applied to optimize variable efficiency and handle missing data. Ensemble algorithms—specifically Gradient Boosting, LightGBM, XGBoost, and Random Forest—achieve high diagnostic discrimination (88–89% accuracy, AUC-ROC > 0.93), outperforming traditional clinical indices.

3. Explainable AI (XAI) Integration To eliminate the "black-box" nature of machine learning and ensure clinical trust, state-of-the-art XAI methods are incorporated:

  • SHAP (Shapley Additive exPlanations): Ranks feature importance globally across the patient cohort, validating that age, CA125, HE4, CEA, and systemic inflammatory or protein markers drive predictions in alignment with known oncological pathophysiology.

  • LIME (Local Interpretable Model-agnostic Explanations): Generates patient-level attributions, showing clinicians the exact weight and direction of individual lab values that contribute to a benign or malignant risk classification.

4. Clinical Significance & Outcomes The model demonstrates the potential to advance ovarian cancer diagnosis (up to 17 days prior to standard pathological confirmation), offering an accessible triage tool that relies primarily on low-cost, routine blood tests rather than complex diagnostic infrastructure. By enabling early identification and immediate routing of high-risk patients to specialized gynecologic oncologists, this interpretable framework supports timely surgical intervention and improved long-term survival rates.

Required knowledge

For IT students conducting research on ovarian cancer prediction, required AI and machine learning knowledge encompasses data preprocessing, modeling, explainable AI, and clinical evaluation: students must understand how to address extreme class imbalance (e.g., SMOTE) and impute missing clinical records, implement ensemble models like LightGBM, XGBoost, and Random Forest that capture non-linear interactions across structured laboratory and demographic features, apply Explainable AI tools like SHAP and LIME to ensure model interpretability for clinicians, and evaluate performance using clinical metrics such as AUC-ROC, precision-recall, and sensitivity rather than simple accuracy alone.

Learn more about minimum entry requirements.

Primary supervisor