Explainable Artificial Intelligence for Ovarian Cancer: Biomarker Contributions in Ensemble Models.

Area of Science:

• Oncology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Ovarian cancer's high mortality necessitates improved early detection.
• Late-stage diagnosis is a primary driver of poor outcomes.
• Existing diagnostic tools require enhancement for timely intervention.

Purpose of the Study:

• To develop and validate explainable artificial intelligence (XAI) models for discriminating malignant from benign ovarian masses.
• To utilize readily available demographic and laboratory data for diagnostic accuracy.
• To provide clinically interpretable insights into diagnostic decision-making.

Main Methods:

• Analysis of a dataset comprising 309 patients (140 malignant, 169 benign) with 47 clinical parameters.
• Feature selection using the Boruta algorithm, identifying 19 significant features including tumor markers, hematological indices, liver function tests, and electrolytes.
• Optimization and evaluation of five ensemble machine learning algorithms via repeated stratified 5-fold cross-validation.

Main Results:

• The Gradient Boosting model demonstrated superior performance with 88.99% accuracy, 0.934 AUC-ROC, and 0.782 Matthews correlation coefficient.
• SHAP analysis highlighted HE4, CEA, globulin, CA125, and age as the most influential features.
• XAI techniques (LIME, SHAP) provided transparent decision pathways, visualizing feature impact on malignancy prediction.

Conclusions:

• Ensemble models using routine lab data can achieve high diagnostic accuracy for ovarian masses, comparable to established clinical indices.
• XAI integration ensures transparency and clinical plausibility, highlighting known and potential novel biomarkers.
• The developed models offer a pragmatic, scalable triage tool to augment ovarian cancer diagnostic pathways, especially in resource-limited settings.