FLOating-Window Projective Separator (FloWPS): A Data Trimming Tool for Support Vector Machines (SVM) to Improve...

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning in Medicine

Background:

• Support Vector Machines (SVMs) are powerful tools for classification but can suffer from extrapolation issues with high-dimensional molecular data.
• Personalized medicine requires accurate prediction models using genetic profiles (e.g., gene expression, mutations) linked to clinical outcomes.
• Existing methods may not effectively handle noisy or irrelevant features in large genetic datasets, limiting classifier performance.

Purpose of the Study:

• To introduce a heuristic data trimming technique, FLOating Window Projective Separator (FloWPS), designed to improve SVM performance for personalized predictions.
• To prevent SVM extrapolation by excluding non-informative features and focusing on clinically relevant genetic profiles.
• To enhance the accuracy and robustness of classifiers built on high-throughput genetic data.

Main Methods:

• Developed FLOating Window Projective Separator (FloWPS), a heuristic method for trimming non-informative features in molecular datasets for SVM.
• FloWPS identifies and removes irrelevant validation dataset features lacking significant neighbors in the training dataset.
• The method adapts the training dataset for each validation point, creating a 'floating window' similar to k-nearest neighbors (kNN).

Main Results:

• FloWPS was tested on ten gene expression datasets from 992 cancer patients undergoing chemotherapy.
• Leave-one-out cross-validation demonstrated that FloWPS significantly increased the quality of SVM classifiers compared to classical SVM.
• Performance improvements were particularly notable for SVMs utilizing polynomial kernels.

Conclusions:

• FloWPS is an effective data trimming strategy that enhances SVM-based personalized prediction models using molecular data.
• The technique addresses feature relevance and prevents extrapolation, leading to more accurate and clinically meaningful classification.
• FloWPS offers a valuable approach for improving predictive accuracy in precision oncology and other fields utilizing high-throughput genetic data.