Rough Set Theory Trims the Data Needed to Flag HER2-Positive Breast Cancer

Deciding whether a breast tumor is HER2-positive is one of the most consequential calls in oncology. Patients whose tumors overexpress the human epidermal growth factor receptor 2 can be treated with targeted antibodies such as trastuzumab, a therapy that has transformed outcomes for this subtype since the early 2000s. But confirming HER2 status normally requires immunohistochemistry and in situ hybridization assays, specialized laboratory infrastructure, and trained pathologists—resources that are far from universally available. A new study published in Neural Computing and Applications asks a deceptively simple question: how far can routinely collected clinical information go toward predicting HER2 status, and can a mathematical framework from the 1980s help strip that information down to its essentials?

The study, conducted by Hoda Waguih of the Sadat Academy for Management Sciences in Cairo, applies Rough Set Theory, a mathematical approach to reasoning about imprecise data introduced by Polish computer scientist Zdzisław Pawlak in 1982, to the problem of feature selection in HER2 classification. Rough set theory occupies an unusual niche in machine learning. Where many feature selection methods rank variables by statistical correlation or by how much a model’s accuracy drops when they are removed, rough sets take a more structural view. The framework partitions patients into groups that are indistinguishable from one another given the available attributes, then identifies which attributes are genuinely needed to separate the decision classes—in this case, HER2-positive versus HER2-negative tumors. Attributes that do not contribute to this separation are redundant and can be discarded without loss of information.

The practical appeal of this approach in a clinical context is hard to overstate. Feature reduction in medical machine learning is often pursued for accuracy, but interpretability and cost matter just as much. A model built on a handful of variables that clinicians already record—demographics, tumor characteristics, routine pathology measures—can be deployed in clinics that lack molecular diagnostics, whereas a model requiring dozens of inputs or proprietary genomic panels cannot. By deriving a minimal subset of features through rough set analysis, the study aims to establish not just a predictive model but a kind of certificate of sufficiency: proof that the retained variables carry essentially all the discriminative information available in the larger set.

The data behind the study come from METABRIC, the Molecular Taxonomy of Breast Cancer International Initiative, a widely used cohort containing clinical profiles of more than 2,500 breast cancer patients. For this analysis, the working dataset comprised 1,294 patients with complete records for the variables of interest. Twelve clinical variables were considered initially, spanning the kind of information gathered in standard oncology workups. The rough set framework reduced this set to eleven variables—a modest trim, but one with an important implication. The near-absence of removable features suggests that the clinically curated set was already highly informative and exhibited low redundancy, meaning nearly every variable earned its place.

Methodological rigor was a stated priority throughout. The author explicitly prevented target leakage, the subtle but common error in which information that would not be available at prediction time contaminates the training data and inflates performance estimates. Model evaluation used stratified five-fold cross-validation, which preserves the class balance across each partition, alongside an independent hold-out test set for final assessment. Statistical significance of performance differences was evaluated with the Wilcoxon signed-rank test at the cross-validation level. These choices matter because clinical machine learning literature is littered with optimistic results that evaporate under leakage-free, properly cross-validated evaluation.

Three classifiers were trained on the reduced feature set: a Decision Tree, a Support Vector Machine with a radial basis function kernel, and XGBoost, a gradient-boosted tree ensemble that has become a mainstay of tabular prediction tasks. Crucially, the models were trained under baseline conditions, without resampling techniques or cost-sensitive learning. This was a deliberate design decision. By keeping the training pipeline minimal, the study isolates the effect of feature reduction itself, ensuring that any performance changes could be attributed to the rough set analysis rather than to compensatory mechanisms layered on top. It also provides an honest picture of how these classifiers behave on an imbalanced clinical problem when nothing is done to correct the imbalance.

The headline result is that rough set-based reduction preserved predictive performance relative to baseline models trained on the full twelve-variable set, with no statistically significant differences observed at the p ≥ 0.05 threshold. Among the three classifiers, the Support Vector Machine achieved the best overall performance, reaching a ROC AUC of 0.72 and a recall of 0.53 for HER2-positive tumors on the hold-out test set. In plain terms, the model ranked patients moderately well overall but caught only about half of the true HER2-positive cases. All three models performed strongly on the majority class—HER2-negative tumors—while showing reduced sensitivity for the positive class, a textbook signature of class imbalance in clinical classification.

That sensitivity gap is where the study is most candid, and most instructive. A recall of 0.53 for the class that triggers targeted therapy is not clinically deployable on its own; missing roughly half of HER2-positive patients would be unacceptable as a screening or triage tool. The author’s framing, however, is not that the models failed but that the experiment reveals something structural about the problem: clinical variables alone, however well curated, carry a ceiling of information about HER2 status, and the biology that ultimately determines receptor overexpression is captured more directly by molecular assays. The value of the exercise lies in quantifying that ceiling honestly, under leakage-free conditions, rather than overstating what routine data can deliver.

The findings also carry a broader lesson about model selection under imbalance. The fact that the SVM outperformed the tree-based models on this task, despite the current enthusiasm for ensemble methods, underscores that algorithm choice interacts with data geometry and class distribution in ways that benchmarks on balanced datasets do not always predict. The author notes that related work, presented at the 2025 ICICIS conference and in a manuscript under review, explores cost-sensitive learning and class imbalance mitigation strategies for the same task, suggesting a research program aimed at pushing the sensitivity of clinical-variable models upward without sacrificing the interpretability that makes them attractive for low-resource settings.

What emerges from the study is a two-part conclusion. First, rough set theory proved most valuable not as an aggressive pruning tool but as a validator: it demonstrated that the eleven-variable clinical set was essentially sufficient, with little redundancy to exploit, which is itself a useful negative result for anyone hoping to shrink clinical feature sets further. Second, the use of routinely available variables supports the development of interpretable, resource-efficient decision-support tools for HER2 classification—tools that could flag patients for priority molecular testing in settings where every assay counts. The path from a ROC AUC of 0.72 to a clinically trusted triage instrument runs through better handling of class imbalance and, ultimately, through integration with rather than replacement of molecular diagnostics. But as a demonstration that transparent mathematics and honest evaluation can map the boundary of what routine clinical data can do, the study offers a template worth copying.

Subject of Research: Interpretable feature reduction using rough set theory for HER2-positive breast cancer classification from clinical data

Article Title: Interpretable feature reduction for HER2-positive breast cancer diagnosis: a rough set theory approach

Article References: Waguih, H. (2026). Interpretable feature reduction for HER2-positive breast cancer diagnosis: a rough set theory approach. Neural Computing and Applications, 38(17), Article 718. https://doi.org/10.1007/s00521-026-12384-6

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12384-6

Keywords: HER2-positive breast cancer, rough set theory, feature selection, interpretable machine learning, METABRIC dataset, support vector machine, XGBoost, class imbalance, clinical decision support, low-resource healthcare, target leakage, cross-validation