Resolving Interpretation Challenges in Machine Learning Feature Selection With an Iterative Approach in Biomedical...

Area of Science:

• Biomedical research
• Computational biology
• Data science

Background:

• Machine learning (ML) is increasingly used for pain data analysis, focusing on classification rather than p-values.
• A challenge exists when accurate classification persists after removing key variables, causing uncertainty about true relevance.
• This ambiguity highlights the need for robust feature selection methods in ML.

Purpose of the Study:

• To present an iterative ML framework for improved identification of trait-relevant features.
• To reduce ambiguity and enhance the interpretability of feature selection in pain research.
• To distinguish robust predictors from coincidental ones in biomedical data.

Main Methods:

• An iterative ML framework was developed, combining feature selection techniques with classification algorithms.
• The framework was applied to pain trait datasets and compared with traditional statistical methods like logistic regression.
• The approach involved repeatedly testing variable groups to assess feature relevance.

Main Results:

• The iterative process clarified variable relevance by testing unselected features.
• Combining ML approaches improved feature selection, addressed multicollinearity, and increased model robustness.
• Logistic regression sometimes failed to identify known relevant variables or required preselected inputs.

Conclusions:

• ML-based feature selection offers expanded options for identifying trait-relevant variables.
• Iterative variable set testing supports transparent and replicable inference.
• Selected features should not be assumed uniquely important; testing unselected variables is crucial.