<p>Effective feature selection is critical for building robust and interpretable predictive models, particularly in medical applications where identifying risk factors in the most extreme patient strata is essential. Traditional methods often focus on average associations, potentially overlooking predictors whose importance is concentrated in the tails of the data distribution. In this study, we introduce a novel, computationally efficient supervised filter that leverages a Gumbel copula implied upper-tail concordance score (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\lambda _U\)</EquationSource> </InlineEquation>, a monotone transformation of Kendall’s <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\tau\)</EquationSource> </InlineEquation>) to rank features by their tendency to be simultaneously extreme with the positive class. We evaluated this method against four standard baselines (Mutual Information, mRMR, ReliefF, and L1/Elastic-Net) across four classifiers on two diabetes datasets: a large-scale public health survey (CDC, <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(N=253,680\)</EquationSource> </InlineEquation>) and a classic clinical benchmark (PIMA, <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(N=768\)</EquationSource> </InlineEquation>). Our analysis included comprehensive statistical tests, permutation importance, and robustness checks. On the CDC dataset, our method was the fastest selector and reduced the feature space by <InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(\approx\)</EquationSource> </InlineEquation>52%. While this resulted in a minimal but statistically significant performance trade-off compared to using all 21 features, our filter significantly outperformed standard filters (Mutual Information, mRMR) and was statistically indistinguishable from the strong ReliefF baseline. On the PIMA dataset (8 predictors), our method’s ranking produced the numerically highest ROC–AUC, despite paired DeLong tests showing no statistically significant differences versus strong baselines. PIMA thus serves as a ranking-only sanity check that our upper-tail criterion behaves sensibly in a low-dimensional clinical setting. Across both datasets, the Gumbel-<InlineEquation ID="IEq6"> <EquationSource Format="TEX">\(\lambda _U\)</EquationSource> </InlineEquation> selector consistently identified clinically coherent and impactful predictors. We conclude that feature selection via upper-tail dependence is an efficient and interpretable screening approach that can complement standard feature-selection baselines in public health and clinical risk prediction.</p>

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

期刊论文
出版模式：

开放获取
发表日期：

2026年3月5日

查看全文

Agnideep Aich,
Md. Monzur Murshed,
Sameera Hewage,
Amanda Mayeaux

摘要

Effective feature selection is critical for building robust and interpretable predictive models, particularly in medical applications where identifying risk factors in the most extreme patient strata is essential. Traditional methods often focus on average associations, potentially overlooking predictors whose importance is concentrated in the tails of the data distribution. In this study, we introduce a novel, computationally efficient supervised filter that leverages a Gumbel copula implied upper-tail concordance score ( \(\lambda _U\) , a monotone transformation of Kendall’s \(\tau\) ) to rank features by their tendency to be simultaneously extreme with the positive class. We evaluated this method against four standard baselines (Mutual Information, mRMR, ReliefF, and L1/Elastic-Net) across four classifiers on two diabetes datasets: a large-scale public health survey (CDC, \(N=253,680\) ) and a classic clinical benchmark (PIMA, \(N=768\) ). Our analysis included comprehensive statistical tests, permutation importance, and robustness checks. On the CDC dataset, our method was the fastest selector and reduced the feature space by \(\approx\) 52%. While this resulted in a minimal but statistically significant performance trade-off compared to using all 21 features, our filter significantly outperformed standard filters (Mutual Information, mRMR) and was statistically indistinguishable from the strong ReliefF baseline. On the PIMA dataset (8 predictors), our method’s ranking produced the numerically highest ROC–AUC, despite paired DeLong tests showing no statistically significant differences versus strong baselines. PIMA thus serves as a ranking-only sanity check that our upper-tail criterion behaves sensibly in a low-dimensional clinical setting. Across both datasets, the Gumbel- \(\lambda _U\) selector consistently identified clinically coherent and impactful predictors. We conclude that feature selection via upper-tail dependence is an efficient and interpretable screening approach that can complement standard feature-selection baselines in public health and clinical risk prediction.