Balancing Performance and Interpretability: Selecting Features with Bootstrapped Ridge Regression.

Area of Science:

• Computational biology
• Health informatics
• Machine learning in healthcare

Background:

• Predictive models for chronic healthcare events often use numerous predictors from diverse datasets.
• This approach enhances performance but reduces model interpretability and portability.
• There is a need for methods that balance predictive accuracy with model understandability.

Purpose of the Study:

• To introduce the Bootstrapped Ridge Selector (BoRidge) as a tool for enhancing predictive model interpretability and performance.
• To compare BoRidge against existing feature selection methods like Bootstrapped LASSO regression (BoLASSO) and minimal-redundancy-maximal-relevance (mRMR).

Main Methods:

• The Bootstrapped Ridge Selector (BoRidge) was developed to balance predictive performance and interpretability.
• BoRidge was evaluated using artificially generated data for binary classification tasks.
• Performance was compared against BoLASSO and mRMR using sensitivity and specificity metrics.
• BoRidge was further validated on a dataset for a published suicide risk prediction model.

Main Results:

• BoRidge demonstrated superior performance in binary classification on artificial data compared to BoLASSO and mRMR (BoRidge: sensitivity 0.83, specificity 0.72).
• On a suicide risk prediction dataset, BoRidge achieved comparable precision to a published model but used significantly fewer predictors (114 vs. 1,538).

Conclusions:

• The Bootstrapped Ridge Selector (BoRidge) offers a promising approach to simplify complex classification models.
• BoRidge enhances model interpretability and portability, facilitating easier translation and actionability in clinical practice.
• This method holds potential for improving the development and application of predictive models in chronic healthcare.