Unsupervised Manifold Learning-Based Hybrid Approaches for Breast Cancer Classification: A Benchmark Study Against Supervised Models

<jats:p>Breast cancer is the most common cancer type among women worldwide, and early diagnosis is critically important for treatment success. Studies in the literature mostly focus on supervised machine learning methods based on labeled medical data. However, this approach may cause human bias to be reflected in the model due to dependence on expert labels and may lead to high labeling costs. In this study, the extent to which unsupervised algorithms can distinguish cancer tissue patterns without human guidance was investigated. Using 569 patient records from the Wisconsin Diagnostic Breast Cancer (WDBC) dataset, a comprehensive repeated benchmark framework was designed with a 10-fold cross-validation and 50-seed stability loop that strictly prevents data leakage. The high-dimensionality problem was addressed with PCA and UMAP; 6 unsupervised hybrid models (K-Means, PAM, GMM, FCM, and their PCA/UMAP-based combinations) and 6 supervised models (Penalized Logistic Regression, SVM, KNN, RF, NB, XGBoost) were tested. For external evaluation, unsupervised clusters were subsequently mapped to the true labels, and cluster quality was validated using the Silhouette, Dunn, and Davies-Bouldin (DBI) indices. The findings showed that the SVM (Linear) model exhibited the highest performance in terms of overall accuracy (0.9658 ± 0.0147); while the Penalized Logistic Regression model stood out in terms of specificity and precision. One noteworthy finding of the study is that the autonomous "UMAP + K-Means" approach, which does not use the cancer label during the modeling stage, reached 0.9453 ± 0.0226 overall accuracy and 0.9445 ± 0.0386 sensitivity levels in new test patients. This model also demonstrated the strongest internal clustering validity, achieving the highest Silhouette score (0.6000 ± 0.0218), the highest Dunn index (0.0548 ± 0.0583) and the lowest Davies-Bouldin value (0.6101 ± 0.0346). In conclusion, while supervised models offer higher overall accuracy in clinical diagnosis, the strong results produced by manifold-supported unsupervised approaches, especially in terms of sensitivity and topological separation, indicate that these methods can be used as an autonomous and reliable "second-opinion" tool in clinical decision support systems.</jats:p>