AI Model Achieves Breakthrough Accuracy in Ovarian Cancer Detection

Traditional tools such as the Risk of Ovarian Malignancy Algorithm (ROMA) can struggle to distinguish between benign and malignant ovarian tumors, especially when biomarker levels fall near diagnostic thresholds. A new artificial intelligence (AI) model now offers a promising solution by delivering strong diagnostic performance while providing clear, interpretable insights into how predictions are made.

The system was developed by researchers at Inonu University (Malatya, Turkey) and Recep Tayyip Erdogan University (Rize, Turkey), using clinical and laboratory datasets drawn from 309 female patients, providing a real-world testbed for improving diagnostic precision. The team built five ensemble machine learning models—Gradient Boosting, CatBoost, XGBoost, LightGBM, and Random Forest—using 47 clinical features, including age, tumor markers (CA125, HE4, CEA, CA19-9, AFP), liver enzymes, blood counts, and electrolytes.

Boruta feature selection identified 19 variables most strongly associated with malignancy risk. Among all tested approaches, the Gradient Boosting model achieved the strongest results, with 88.99% accuracy, an AUC-ROC score of 0.934, and an MCC of 0.782. This performance surpassed the ROMA index, which achieved an AUC of 0.89 on the same dataset. At 90% specificity, the model detected 82% of malignant cases compared with ROMA’s 78%.

The system incorporates explainable AI frameworks to ensure transparency. Using SHAP and LIME, the researchers could interpret model outputs at both global and individual patient levels. SHAP analysis identified HE4, CEA, globulin (GLO), CA125, and age as the most influential predictors of malignancy, aligning with known oncologic evidence. LIME provided patient-specific explanations, showing how elevated or normal biomarker values shifted the prediction toward malignant or benign categories. These interpretable layers help clinicians validate the model’s reasoning and support adoption in routine diagnostic pathways.

The findings, published in Biology, also showed how error patterns can guide safer clinical application. Most false negatives occurred among patients with biomarker values near borderline ranges or atypical profiles, emphasizing the need for cautious threshold setting to avoid missed cancers. Although the dataset came from a single medical center, the authors recommend larger, multi-center validation before clinical rollout.

The study highlights how explainable AI can enhance cancer screening using non-invasive, cost-effective data already available in clinical settings. By strengthening early risk assessment, especially in resource-limited environments, such tools could improve ovarian cancer outcomes and support more equitable access to diagnostic care. The authors suggest that future systems may combine AI predictions with imaging and genomic data to further refine accuracy and personalize treatment strategies.

Related Links:
Inonu University
Recep Tayyip Erdogan University