Comparison of variable selection methods for clinical predictive modeling.

Area of Science:

• Clinical informatics
• Machine learning in healthcare
• Predictive modeling

Background:

• Machine learning is increasingly used for clinical problems, particularly in variable selection for predictive models.
• Limited research exists comparing classic and modern variable selection methods in clinical datasets.

Purpose of the Study:

• To compare the performance of classic regression-based and modern tree-based variable selection methods on clinical datasets of varying sizes.
• To evaluate how dataset size impacts the effectiveness of different variable selection techniques.

Main Methods:

• Analyzed eight variable selection methods: four regression-based (stepwise backward selection, LASSO, Elastic Net) and four tree-based (Variable Selection Using Random Forest, Regularized Random Forests, Boruta, Gradient Boosted Feature Selection).
• Utilized two clinical datasets: a large multicenter adult cohort and a smaller single-center pediatric cohort.
• Evaluated methods based on parsimony, variable importance, and discrimination.

Main Results:

• Modern tree-based methods (Variable Selection Using Random Forest, Gradient Boosted Feature Selection) showed best parsimony on the large dataset.
• Classic regression-based methods (stepwise backward selection) achieved best parsimony on the smaller dataset.
• Variable selection generally decreased Random Forest model accuracy but increased logistic regression model accuracy.

Conclusions:

• The performance of variable selection methods is dataset-size dependent.
• Classic regression methods are suitable for smaller clinical datasets, while modern tree-based methods are better for larger datasets.
• Choosing the right variable selection method is crucial for effective clinical predictive modeling.