Shift-adjusted Neyman-Pearson classifiers via single index modeling (SACSIM)

Area of Science:

• Statistical Learning
• Machine Learning for Healthcare
• Biostatistics

Background:

• Neyman-Pearson (NP) classifiers are vital for maximizing clinical benefit under risk constraints, particularly in early cancer detection.
• Discrepancies in data distributions between source and target populations hinder NP classifier generalizability, disproportionately affecting under-represented groups.
• Existing methods struggle to adapt NP classifiers equitably across different populations.

Purpose of the Study:

• To develop a semi-parametric model-based approach for adapting NP classifier decision rules to new populations.
• To ensure equitable control of classification errors specific to clinical applications.
• To enhance the applicability and consistency of NP classifiers across diverse demographic groups.

Main Methods:

• A shift-adjustment strategy is employed, utilizing labeled source data and a small unlabeled target sample.
• Minimal auxiliary information is incorporated to bridge the distribution gap.
• The approach focuses on adapting decision rules rather than retraining the entire model.

Main Results:

• Theoretical studies and simulations demonstrate the effectiveness of the proposed adaptation method.
• The approach successfully tailors NP classifier decision rules for target populations.
• Equitable control of classification errors is achieved, even with distribution shifts.

Conclusions:

• The proposed semi-parametric approach offers a robust solution for adapting NP classifiers across populations.
• This method enhances fairness and reliability in high-stakes clinical applications like cancer screening.
• The strategy provides a practical framework for deploying NP classifiers in real-world, diverse settings.