Classifying Incomplete Gene-Expression Data: Ensemble Learning with Non-Pre-Imputation Feature Filtering and...

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Gene-expression datasets frequently contain missing values (MVs).
• While MV imputation methods are abundant, their impact on downstream classification performance is minimal.
• Prioritizing informative gene selection over MV imputation is crucial for classification tasks.

Purpose of the Study:

• To develop a feature selection (FS) method that directly processes incomplete gene-expression data without imputation.
• To evaluate the impact of MV imputation on downstream FS performance.
• To identify informative genes for classification from gene-expression data with missing values.

Main Methods:

• A modified chi-square test-based FS approach is proposed for gene-expression data.
• Recursive element aggregation is introduced to address small sample sizes in gene-expression data.
• The method directly handles incomplete data, avoiding imputation and its assumptions, followed by best-first search for optimal feature subsets.

Main Results:

• The proposed method was compared with existing FS algorithms on twelve incomplete cancer gene-expression datasets.
• MV imputation can negatively impact subsequent FS and classification performance.
• Directly applying FS to incomplete data avoids imputation-related disturbances, leading to potentially better feature subsets.

Conclusions:

• The developed FS method effectively handles missing values in gene-expression data without imputation.
• This approach offers a more robust alternative to traditional FS methods that require prior imputation.
• The method demonstrated improved gene discovery by identifying additional informative genes on the SRBCT dataset.