Optimal clustering with missing values.

Area of Science:

• Biostatistics
• Computational Biology
• Genomics

Background:

• Missing values are common in biomedical studies, complicating clustering.
• Imputation is a standard but potentially flawed approach to handle missing data before clustering.

Purpose of the Study:

• To develop an optimal clustering framework that directly addresses missing values.
• To integrate missing data mechanisms into the random labeled point process (RLPP) for robust clustering.

Main Methods:

• Incorporating missing value mechanisms into the RLPP framework.
• Marginalizing out the missing-value process within optimal clustering.
• Demonstration using Gaussian models with arbitrary covariance structures.

Main Results:

• The proposed optimal clustering framework effectively handles missing values without imputation.
• Experimental studies on synthetic and RNA-seq data show superior performance compared to existing methods.
• The approach achieves smaller clustering errors in the presence of missing data.

Conclusions:

• Optimal clustering with missing values eliminates the need for imputation pre-processing.
• This method offers improved accuracy and reduced clustering errors for biomedical data with missing values.