Diagnostic biases in translational bioinformatics.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Complex disease diagnosis increasingly relies on omics data analysis.
• Translational bioinformatics faces challenges in detecting and preventing diagnostic biases.
• Ensuring accuracy is crucial for the future of personalized medicine.

Purpose of the Study:

• To comprehensively investigate diagnostic bias in omics data analysis.
• To categorize and understand different types of diagnostic biases.
• To develop effective machine learning methods to overcome these biases.

Main Methods:

• Analysis of benchmark gene array, protein array, RNA-Seq, and miRNA-Seq data.
• Application of support vector machines (SVM) with various model selection methods.
• Rigorous kernel matrix analysis to categorize diagnostic biases.
• Development of derivative component analysis based SVM (DCA-SVM).

Main Results:

• Diagnostic biases occur across different data distributions and SVM kernels.
• Identified three types of biases: overfitting, label skewness, and underfitting.
• Label skewness bias is particularly challenging due to deceptive accuracy.
• DCA-SVM demonstrated effective bias correction, achieving competitive clinical diagnostic results.

Conclusions:

• Diagnostic biases stem from kernel selection, signal amplification, and training data label distribution.
• DCA-SVM offers a robust solution for label skewness bias via enhanced feature extraction.
• This research addresses a critical, under-explored issue in translational research.
• The findings contribute to kernel-based learning for omics data analysis.