Ontology-Enhanced Machine Learning Models for Breast Cancer Diagnosis

Abstract

INTRODUCTION: Breast cancer remains one of the most prevalent causes of cancer-related mortality among women globally. While machine learning (ML) has demonstrated promise in early detection, conventional models often rely solely on statistical features, lacking domain-specific knowledge and interpretability.

OBJECTIVES: This study aims to enhance breast cancer prediction by integrating ontology-driven semantic features with ML models to improve both predictive accuracy and clinical interpretability.

METHODS: We applied a comprehensive pipeline comprising data preprocessing, statistical testing, and dimensionality reduction using PCA, followed by training with supervised learning models including Logistic Regression, k-NN, SVM, Random Forest, XGBoost, LightGBM, and Attention-Enhanced MLP. In the proposed approach, clinical data is transformed into RDF triples and structured within a domain-specific breast cancer ontology. Semantic reasoning via SPARQL queries enables the extraction of high-level features, which are then used in a leakage-safe stacking design that integrates (i) tabular features, (ii) KGE features, (iii) semantic subtyping signals, and (iv) SPARQL rule features, with reproducible templates and released code.

RESULTS: Across four benchmark datasets, the ontology-enhanced meta-learner achieved consistently strong performance, achieving 0.996 ± 0.006 ROC-AUC on WDBC under stratified evaluation.

CONCLUSION: Incorporating ontology-derived semantic knowledge significantly improves the performance, robustness, and interpretability of ML models for breast cancer prediction. This approach holds strong potential for real-world integration into clinical decision support systems.