Selecting features with group-sparse nonnegative supervised canonical correlation analysis: multimodal prostate...

Area of Science:

• Biomedical data analysis
• Machine learning for bioinformatics
• Computational pathology

Background:

• Existing correlation-based feature selection methods often fail to guarantee positive feature correlations and require pre-processing for feature redundancy.
• There is a need for methods that simultaneously perform feature selection within and between multiple data views while ensuring positive correlations.

Purpose of the Study:

• To introduce Group-sparse Nonnegative supervised Canonical Correlation Analysis (GNCCA), a novel methodology for identifying discriminative features from multiple data views.
• To address limitations of existing methods by incorporating nonnegativity and group-sparsity constraints for improved feature selection and class separability.

Main Methods:

• Developed GNCCA, integrating nonnegativity constraints for positive correlations and group-sparsity for simultaneous feature selection across and within views.
• Emphasized correlations between feature views and class labels to enhance class separability.
• Applied GNCCA to three prostate cancer (CaP) prognosis tasks, including recurrence prediction, grade prediction, and region localization using diverse data types (pathology, proteomics, MRI, MR spectroscopy).

Main Results:

• GNCCA identified highly reduced feature subsets (2%, 1%, and 22%) for the three CaP tasks.
• The selected features, when used with a Support Vector Machine (SVM) classifier, achieved improved or comparable performance to using all features.
• GNCCA consistently outperformed five state-of-the-art feature selection methods across all evaluated datasets.

Conclusions:

• GNCCA is an effective method for discriminative feature identification from multiple data views, particularly in complex biomedical applications like prostate cancer prognosis.
• The nonnegativity and group-sparsity constraints enable robust feature selection, leading to enhanced predictive performance and improved class separability.
• GNCCA offers a significant advancement over existing feature selection techniques, demonstrating its utility and superiority in real-world datasets.