Adaptive Dimensionality Reduction with Semi-Supervision (AdDReSS): Classifying Multi-Attribute Biomedical Data.

Area of Science:

• Biomedical data analysis
• Machine learning in healthcare
• High-dimensional data processing

Background:

• Medical diagnostics involves complex, high-dimensional biomedical data analysis.
• Key challenges include high feature dimensionality and limited labeled training data.
• Existing methods lack active learning integration for dimensionality reduction.

Purpose of the Study:

• To introduce AdDReSS (Adaptive Dimensionality Reduction with Semi-Supervision), a novel methodology.
• To demonstrate AdDReSS's effectiveness in creating discriminative low-dimensional representations using active learning.
• To improve classification accuracy in high-dimensional biomedical datasets with reduced labeling effort.

Main Methods:

• AdDReSS integrates active learning (AL) with semi-supervised dimensionality reduction.
• The method identifies informative labeled instances in embedding space.
• Tested across diverse biomedical domains: gene expression, proteomics, MRI, and histopathology.

Main Results:

• AdDReSS achieved a median classification accuracy of 88.7%, outperforming SSAGE (85.5%) and Graph Embedding (81.5%).
• Embeddings generated by AdDReSS showed a 35.95% improvement in Raghavan efficiency over SSAGE.
• Fewer labeled examples were required with AdDReSS for comparable or superior performance.

Conclusions:

• AdDReSS provides effective low-dimensional representations for high-dimensional biomedical data.
• The methodology enhances classification rates while minimizing the need for labeled data.
• AdDReSS offers a valuable advancement for medical diagnostics and biomedical research.