Sparse robust discriminant analysis for high-dimensional and heavy-tailed data.

Area of Science:

• Medical data analysis
• Statistical learning
• Bioinformatics

Background:

• Large-scale medical data (gene expression, MRI) are prevalent.
• Existing sparse discriminant analysis methods assume light-tailed predictors, often violated in practice.
• Robustness is needed for heavy-tailed medical data.

Purpose of the Study:

• Propose a robust classifier using an elliptically contoured discriminant analysis (EDA) model.
• Accommodate both light-tailed and heavy-tailed data distributions.
• Improve prediction accuracy on imbalanced medical datasets.

Main Methods:

• Developed a robust classifier under the EDA model.
• Identified intrinsic dimension-reduction subspace for optimal prediction.
• Proposed a high-dimensional classifier using subspace projection.
• Utilized balanced rate for assessing prediction accuracy on imbalanced data.

Main Results:

• The proposed EDA-based classifier accommodates heavy-tailed data.
• Achieved superior prediction accuracy, particularly for imbalanced datasets.
• Demonstrated consistency in subspace estimation, variable selection, and prediction accuracy.
• Empirical results on synthetic and real medical data (lung cancer, leukemia) show superiority over state-of-the-art methods.

Conclusions:

• The proposed robust EDA classifier effectively handles high-dimensional, potentially heavy-tailed medical data.
• Subspace identification and projection offer a powerful approach for robust classification.
• The method provides a more accurate and reliable tool for medical data analysis.