Machine Learning Strategies for Improved Phenotype Prediction in Underrepresented Populations

Area of Science:

• Genomic Medicine
• Bioinformatics
• Machine Learning

Background:

• Precision medicine models exhibit performance disparities due to over-representation of European ancestry in genomic datasets.
• This imbalance leads to less accurate predictions and treatment recommendations for underrepresented populations, exacerbating health inequities.
• Existing genomic prediction models require enhancement to ensure equitable application across diverse ancestries.

Approach:

• Developed an adaptable machine learning toolkit integrating gradient boosting and automated methods.
• Incorporated novel population-conditional re-sampling techniques to address dataset diversity.
• Evaluated the toolkit using the UK Biobank, including diverse Asian and African ancestry groups.

Key Points:

• The toolkit significantly improves phenotypic prediction from single nucleotide polymorphism (SNP) data for diverse populations.
• Achieved prediction accuracy for underrepresented groups comparable to the majority European ancestry group.
• Demonstrated substantial improvements in phenotype prediction for minority groups within the UK Biobank.

Conclusions:

• The developed approach enhances the accuracy and equity of genomic prediction models.
• Represents a significant advancement in addressing dataset diversity challenges in statistical genetics.
• Fosters more inclusive models and equitable outcomes in precision medicine for all populations.