Abstract
BACKGROUND: Machine learning models in biomedical research are often hindered by demographic imbalances in clinical datasets, leading to biased predictions that disadvantage minority populations. Existing bias-correction methods face limitations in handling the heterogeneity of biomedical data and the complexity of demographic influences. RESULTS: We present DeBias, a computational framework for mitigating demographic biases in high-dimensional biomedical datasets. DeBias identifies and removes bias-associated subspaces from the feature space using control samples, enabling global correction of demographic distortions while preserving disease-specific signals. To evaluate its effectiveness, we apply DeBias to cell-free DNA methylation data for cancer detection. DeBias achieves a significant reduction in the number of features exhibiting demographic bias and outperforms existing methods in improving cancer detection performance for minority populations. Performance gains are validated in independent cohorts, highlighting the robustness of the approach. CONCLUSIONS: DeBias offers an effective and generalizable strategy for correcting demographic biases in biomedical machine learning. It represents a step toward more equitable machine learning models that can deliver reliable and unbiased predictions across diverse patient populations.