Ensemble feature selection with data-driven thresholding for Alzheimer's disease biomarker discovery

Area of Science:

• Computational Biology
• Bioinformatics
• Neuroscience

Background:

• High-dimensional data often leads to unstable feature selection results.
• Ensemble methods enhance feature selection stability by aggregating multiple base selectors.
• Fixed thresholds in ensemble methods may not guarantee relevant feature identification.

Purpose of the Study:

• To evaluate data-driven thresholds for automatic relevant feature identification in ensemble feature selection.
• To assess the predictive accuracy and stability of ensemble feature selection with data-driven thresholds.
• To apply these methods to Alzheimer's disease (AD) datasets for biomarker discovery.

Main Methods:

• Ensemble feature selection incorporating data-driven thresholds.
• Evaluation of robust rank aggregation and information retrieval threshold algorithms.
• Application to two real-world Alzheimer's disease datasets.

Main Results:

• Ensemble methods with data-driven thresholds improved stability by up to 34% compared to individual selectors.
• Robust rank aggregation and threshold algorithm proved most effective.
• Identified features align with current Alzheimer's disease literature.

Conclusions:

• Data-driven thresholds enhance feature selection stability and reproducibility in ensemble methods.
• Eliminates the need for arbitrary fixed thresholds, leading to more meaningful feature sets.
• Enables the development of interpretable models by highlighting key disease factors.