Feature Selection in Healthcare Datasets: Towards a Generalizable Solution.

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Data Science

Background:

• High-dimensional healthcare datasets pose challenges for clinical data analysis and interpretation.
• Dimensionality reduction is crucial for efficient and accurate clinical insights.
• Current methods may struggle with multi-biometric and heterogeneous data.

Purpose of the Study:

• To develop a scalable ensemble feature selection (FS) strategy for multi-biometric healthcare datasets.
• To address dimensionality reduction needs and identify significant clinical features.
• To enhance machine learning model performance and clinical interpretability.

Main Methods:

• A novel waterfall selection integrating tree-based feature ranking and greedy backward elimination.
• A merging strategy to combine feature subsets into a single, clinically relevant set.
• Application to diverse datasets: BioVRSea (biosignals) and SinPain (medical images).

Main Results:

• Achieved significant dimensionality reduction, exceeding 50% in some feature subsets.
• The reduced feature set maintained or improved classification metrics (Support Vector Machine, Random Forest).
• Demonstrated effectiveness on both biosignal and image-based healthcare data.

Conclusions:

• The ensemble FS method retains essential features for clinical outcome discrimination.
• Resulting models are computationally efficient and clinically interpretable.
• The scalable and adaptable approach shows potential as a generalizable tool for healthcare studies.