Predictive and interpretable models via the stacked elastic net.

Area of Science:

• Biomedical Sciences
• Machine Learning
• Statistical Modeling

Background:

• Machine learning models in biomedical sciences need to be both predictive and interpretable.
• Researchers require models that identify feature effects (direction and strength) for clinical and molecular data.
• Traditional regression offers interpretability but lacks the predictive power of advanced machine learning.

Purpose of the Study:

• To propose an interpretable meta-learning approach for high-dimensional regression.
• To enhance the predictive performance of machine learning models in biomedical applications.
• To maintain model interpretability alongside increased predictivity.

Main Methods:

• Developed an interpretable meta-learning framework for high-dimensional regression.
• Utilized the elastic net, which balances regularization between L1 (Lasso) and L2 (Ridge) penalties.
• Employed a stacking technique to combine multiple elastic net weightings, bypassing single-parameter tuning.

Main Results:

• The proposed meta-learning approach increases model predictivity.
• Interpretability is preserved, allowing for the identification of feature effects.
• The method effectively handles high-dimensional data common in biomedical research.

Conclusions:

• The novel meta-learning strategy offers a powerful tool for biomedical researchers.
• It bridges the gap between predictive accuracy and model interpretability in regression tasks.
• The R package 'starnet' is available for practical implementation.