Beyond P-values: A Multi-Metric Framework for Robust Feature Selection and Predictive Modeling.

Area of Science:

• Biomedical data analysis
• Machine learning in healthcare
• Genomics and multi-omics

Background:

• High-dimensional biomedical datasets often have sparse signals within correlated features.
• Variable selection is crucial for developing generalizable predictive models.
• Current methods often focus on statistical significance, which doesn't always ensure predictive utility.

Purpose of the Study:

• To develop a domain-agnostic approach for integrating multiple variable selection criteria.
• To create a unified framework that combines inferential and predictive evidence for feature selection.
• To improve the accuracy, interpretability, and transportability of predictive models in biomedical research.

Main Methods:

• Introduction of MIXER (Multi-metric Integration for eXplanatory and prEdictive Ranking).
• Adaptive weighting to integrate multiple selection metrics into a consensus model.
• Simulation studies to evaluate feature set overlaps based on data characteristics.
• Application to Alzheimer's disease data from UK Biobank and external validation.

Main Results:

• Simulation studies showed that different selection metrics identify distinct feature sets.
• MIXER outperformed individual selection criteria, including statistical significance, in Alzheimer's disease prediction.
• The model generalized to an external cohort (Alzheimer's Disease Sequencing Project), showing improved discrimination and risk stratification.
• Demonstrated the modularity and extensibility of the MIXER framework.

Conclusions:

• MIXER provides a practical approach to enhance variable selection in high-dimensional biomedical data.
• Integrating multiple metrics leads to more robust and reliable predictive models.
• The framework offers a pathway to more accurate, interpretable, and transportable biomedical predictive models.