A generic model-free feature screening procedure for ultra-high dimensional data with categorical response

Area of Science:

• Statistical Learning
• Bioinformatics
• Genomics

Background:

• Identifying active features is crucial in ultra-high dimensional data analysis.
• This task is vital for statistical learning and biological discovery.

Purpose of the Study:

• To develop a novel procedure for feature selection in ultra-high dimensional data with categorical responses.
• To address challenges in biomedical studies like category-adaptive data and unbalanced distributions.

Main Methods:

• A model-free, nonparametric concordance index screening (CI-SIS) procedure was developed.
• The method utilizes the concordance index measure for sure screening and ranking consistency.
• A data-driven threshold selection via knockoff features was also presented.

Main Results:

• CI-SIS demonstrated lower prediction error on a lung dataset (0.107 with LDA, 0.117 with RF).
• It achieved accuracy improvements of 3% (LDA) and 5% (RF) over the runner-up.
• On SRBCT data, CI-SIS showed at least an 8% performance improvement compared to other methods.

Conclusions:

• The proposed method efficiently identifies disease-associated genes.
• Selected features aid in precision diagnosis and refined patient treatment.