Reducing demographic bias in biomedical machine learning for cancer detection using cfDNA methylation

Area of Science:

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

Background:

• Demographic imbalances in clinical datasets cause biased machine learning predictions, disadvantaging minority populations.
• Current bias-correction methods struggle with biomedical data heterogeneity and complex demographic factors.

Purpose of the Study:

• To introduce DeBias, a novel computational framework designed to mitigate demographic biases in high-dimensional biomedical datasets.
• To enhance the fairness and accuracy of machine learning models in biomedical research.

Main Methods:

• DeBias identifies and removes bias-associated subspaces from feature spaces using control samples.
• The framework enables global correction of demographic distortions while preserving crucial disease-specific signals.
• Applied to cell-free DNA methylation data for cancer detection.

Main Results:

• DeBias significantly reduced features exhibiting demographic bias.
• Outperformed existing methods in improving cancer detection performance for minority populations.
• Performance gains were validated in independent cohorts, demonstrating robustness.

Conclusions:

• DeBias provides an effective and generalizable strategy for correcting demographic biases in biomedical machine learning.
• Represents progress towards equitable machine learning models for reliable, unbiased predictions across diverse patient groups.