2D-EM clustering approach for high-dimensional data through folding feature vectors.

Area of Science:

• Biomedical data analysis
• Computational biology
• Machine learning in healthcare

Background:

• Biomedical research generates large, complex datasets.
• Unsupervised clustering identifies patient phenotypes but struggles with high dimensionality and low sample sizes.
• Existing algorithms fail to accurately cluster biological data with many features and few samples.

Purpose of the Study:

• To develop a novel clustering algorithm for high-dimensional, low-sample biomedical data.
• To address the limitations of current clustering methods in biological data analysis.

Main Methods:

• Introduced 2D-EM, a clustering algorithm for small sample size, high-dimensional datasets.
• Transformed data into a two-dimensional matrix for visualization and distribution analysis.
• Utilized a modified expectation-maximization (EM) algorithm to estimate maximum likelihood.

Main Results:

• 2D-EM demonstrated superior performance on transcriptome and methylome datasets.
• Achieved up to 21.9% improvement in Rand score and 155.6% in adjusted Rand index compared to existing methods.
• Successfully reproduced known groups in transcriptome and methylome data with high accuracy.

Conclusions:

• The 2D-EM algorithm accurately clusters challenging biomedical datasets, outperforming established methods.
• Its design supports diverse datasets and enhances the discovery of novel disease subtypes.
• A MATLAB implementation is available online for public use.