A copula based supervised filter for feature selection in machine learning driven diabetes risk prediction

Area of Science:

• Machine Learning
• Biostatistics
• Medical Informatics

Background:

• Effective feature selection is crucial for interpretable predictive models in medicine.
• Traditional methods may miss predictors important in extreme patient groups.
• Identifying risk factors in the tails of data distributions is essential for targeted interventions.

Purpose of the Study:

• To introduce a novel, computationally efficient supervised filter for feature selection.
• To rank features based on their tendency to be extreme with the positive class using Gumbel copula upper-tail concordance.
• To evaluate the proposed method against standard baselines on diabetes datasets.

Main Methods:

• Developed a supervised filter leveraging Gumbel copula implied upper-tail concordance score.
• Evaluated the filter against Mutual Information, mRMR, ReliefF, and L1/Elastic-Net.
• Tested on two diabetes datasets (CDC and PIMA) using four classifiers.
• Conducted statistical tests, permutation importance, and robustness checks.

Main Results:

• The Gumbel-based selector was the fastest on the CDC dataset, reducing features by ≈52% with minimal performance trade-off.
• It significantly outperformed Mutual Information and mRMR, and was comparable to ReliefF.
• On the PIMA dataset, the ranking yielded the highest ROC-AUC numerically.
• The method consistently identified clinically relevant predictors across both datasets.

Conclusions:

• Feature selection via upper-tail dependence is an efficient and interpretable screening approach.
• This method can effectively complement standard feature selection techniques.
• It is particularly valuable for public health and clinical risk prediction, focusing on extreme patient strata.