Omics-based Cancer Prognosis Across Ethnic Groups: From Feature Engineering to Disparity Detection and Mitigation.

Area of Science:

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

Background:

• Artificial Intelligence (AI) models face challenges with unequal ethnic representation in biomedical data, leading to data inequality and poor generalization.
• High-dimensional omics data further complicates AI model analysis, particularly for underrepresented or data-disadvantaged groups.

Purpose of the Study:

• To develop and validate a machine learning approach for omics-based cancer prognosis that addresses data inequality across ethnic groups.
• To detect and mitigate performance disparities in AI models caused by imbalanced biomedical datasets.
• To improve AI model outcomes for underrepresented populations in precision medicine.

Main Methods:

• An autoencoder-based feature engineering method was employed for feature extraction from omics data.
• Machine learning schemes including mixture learning, independent learning, and transfer learning were applied across ethnic groups.
• Transfer learning was utilized to mitigate identified performance gaps for data-disadvantaged groups.

Main Results:

• Experimental analysis revealed performance disparities in 22.34% of machine learning tasks.
• Transfer learning effectively mitigated disparities for the African American group using mRNA Expression features from The Cancer Genome Atlas dataset.
• The approach successfully identified and reduced disparities in DNA methylation and MicroRNA Expression omics datasets for data-disadvantaged groups.

Conclusions:

• The presented approach demonstrates efficacy in detecting performance disparities in AI models due to biomedical data inequality.
• Transfer learning significantly enhances AI model performance and mitigates the effects of data disparities across diverse ethnic groups.
• This method offers a pathway to more equitable and accurate AI-driven cancer prognosis.