Predicting breast cancer using an expression values weighted clinical classifier

Minta Thomas, Kris De Brabanter, Johan A. K. Suykens, Bart De MoorBMC Bioinformatics 2014 · DOI

Introduction

In modern oncology, clinicians rely on two distinct types of data for prognosis and treatment planning: traditional clinical parameters—such as age, tumor size, and nodal status—and high-dimensional microarray data, which captures the expression levels of thousands of genes. While clinical variables are well-established and reliable, genomic data offers the potential for a more granular understanding of a patient's unique biological profile.

The primary challenge in this field is "data fusion." Microarray data is characterized by high dimensionality and a low signal-to-noise ratio, which often leads to overfitting or the drowning out of reliable clinical signals by genomic "noise." Previous attempts to integrate these datasets have produced conflicting results, with some studies showing synergy and others suggesting that genomic data adds little value over clinical variables alone. This research introduces a robust mathematical framework to demonstrate that effective integration is possible through a weighted machine learning approach.

Methodology

The core of this research is the development of a weighted Least Squares Support Vector Machine (LS-SVM) classifier. This approach is built upon three fundamental pillars:

  1. Least Squares Support Vector Machines (LS-SVM): A computationally efficient variant of standard SVMs that solves linear equations rather than complex quadratic programming problems, making it ideal for high-dimensional genomic data.
  2. The Kernel Trick: To handle heterogeneous data types (e.g., continuous age vs. categorical ER status), the researchers utilized kernel functions. This transforms diverse data structures into a unified "kernel matrix," allowing for seamless mathematical combination.
  3. Weighted Integration: Unlike traditional methods that simply concatenate features, this model uses microarray data to generate a weighting matrix. This matrix "tunes" the clinical classifier, allowing genomic information to emphasize or de-emphasize specific clinical patterns based on the molecular context.
Parameter optimization algorithm of weighted LS-SVM classifier

To ensure the model generalizes well to unseen patients, hyperparameters—including the regularization constant $\gamma$ and kernel bandwidths $\sigma$—are optimized using Leave-One-Out Cross-Validation (LOO-CV), as illustrated in the optimization workflow above.

Datasets and Experimental Setup

The model was rigorously evaluated using five distinct breast cancer case studies (Cases I–V) from the Integrated Tumor Transcriptome Array and Clinical data Analysis (ITTACA) database. These datasets featured:

  • Sample Sizes: 85 to 251 patients.
  • Genomic Complexity: Microarray data encompassing between 5,000 and 20,055 genes.
  • Clinical Depth: Standard metrics including ethnicity, ER/PR status, tumor grade, and therapeutic response.

Results and Discussion

The weighted LS-SVM consistently outperformed individual datasets and alternative fusion techniques, such as Generalized Eigenvalue Decomposition (GEVD), across all five case studies. The primary performance metric was the Area Under the ROC Curve (AUC).

Comparison of the prediction accuracy of the classifiers across five breast cancer cases

Key performance highlights include:

  • Case I: Achieved an AUC of 0.8177, a significant improvement over clinical data alone (0.7795) or microarray data alone (0.7001).
  • Case II: Reached an AUC of 0.8465.
  • Case V: Reached an AUC of 0.8210.

In nearly every instance, the weighted approach provided a statistically significant improvement (p < 0.05) over isolated data types. These results suggest that the method of fusion is as critical as the data itself; by treating microarray data as a contextual weighting mechanism rather than just additional features, the model respects the primary importance of clinical indicators while leveraging the nuance of genomics.

Conclusion and Future Directions

This study demonstrates that integrating high-dimensional genomic data with clinical parameters can significantly enhance breast cancer prediction accuracy. However, the authors note that the current optimization procedure is computationally intensive due to the simultaneous tuning of multiple parameters via LOO-CV.

Future research will focus on improving computational efficiency, applying this weighted framework to other "omics" data (such as proteomics), and developing real-time clinical decision support tools to assist oncologists in personalized therapy selection.

Summary updated

BibTeX
@article{thomas2014predicting,
  author = {Minta Thomas and Kris De Brabanter and Johan A. K. Suykens and Bart De Moor},
  title = {{Predicting breast cancer using an expression values weighted clinical classifier}},
  journal = {BMC Bioinformatics},
  year = {2014},
  doi = {10.1186/s12859-014-0411-1},
}
Figure