Subpopulation-specific machine learning prognosis for underrepresented patients with double prioritized bias...

Area of Science:

• Machine Learning
• Clinical Prognosis
• Health Disparities

Background:

• Clinical datasets often exhibit imbalance, with majority groups dominating.
• Standard machine learning models may fail minority patient groups and demographics, leading to disparities.
• Existing whole-population metrics can be misleading, masking these biases.

Purpose of the Study:

• To develop and evaluate a novel bias correction technique for machine learning-based prognosis.
• To address representational biases affecting minority patient groups in clinical predictions.
• To compare the proposed method against existing bias mitigation strategies.

Main Methods:

• Introduced a double prioritized (DP) bias correction technique.
• Developed customized machine learning models for specific ethnicity or age groups.
• Compared DP with sampling and reweighting techniques on mortality and cancer survivability prediction tasks.

Main Results:

• Demonstrated significant prediction deficiencies in standard machine learning without bias correction.
• DP consistently improved minority class recall by up to 38.0%.
• DP reduced relative disparities across race and age groups, outperforming existing methods by up to 88.0%.

Conclusions:

• The one-model-fits-all approach in machine learning perpetuates biases.
• A novel bias correction method creates specialized models for underrepresented groups.
• This technique can decrease critical prediction errors for minority populations.