Benchmarking feature selection and feature extraction methods to improve the performances of machine-learning...

Area of Science:

• Biomedical data analysis
• Machine learning in biology

Background:

• Omics data present challenges for machine learning due to heterogeneity, sparsity, and high dimensionality (p >> n).
• Class and feature imbalance are common issues in multi-omics datasets, hindering accurate classification.
• Metabolomics is crucial for understanding cancer, but its data complexity requires advanced analytical methods.

Purpose of the Study:

• To investigate the impact of feature extraction and selection techniques on machine learning classification performance with omics data.
• To develop guidelines for applying these techniques based on data characteristics.
• To evaluate the generalizability of these methods across different omics types.

Main Methods:

• Applied various linear and non-linear feature extraction techniques to three metabolomics datasets.
• Evaluated feature extraction methods with and without supervised feature selection.
• Tested the approach on transcriptomics and proteomics datasets to assess broader applicability.

Main Results:

• Supervised feature selection consistently improved the classification performance of feature extraction methods across all tested omics datasets.
• Demonstrated improved patient classification accuracy using the proposed feature engineering strategies.
• Provided a generalizable workflow and guidelines for optimizing omics data classification.

Conclusions:

• Feature selection is a critical step for enhancing machine learning model performance in omics data analysis.
• The findings offer practical strategies for improving classification accuracy in cancer research and other biomedical fields.
• The study provides valuable insights and tools for researchers working with complex biological datasets.