Predictive models for low birth weight: a comparative analysis of algorithmic fairness-improving approaches.

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Health Equity Research

Background:

• Predictive models for low birth weight are crucial for population health management.
• Algorithmic fairness is essential to ensure equitable health outcomes.
• Existing models may exhibit performance disparities across racial/ethnic groups.

Purpose of the Study:

• To evaluate if common algorithmic fairness-improving approaches enhance predictive models for low birth weight.
• To assess the impact of fairness interventions on model accuracy and sensitivity across different racial/ethnic groups.

Main Methods:

• Retrospective analysis of linked birth certificates and insurance claims (n=191,943).
• Comparison of an original elastic net model with 6 fairness-improving approaches.
• Model training and testing using balanced random splits and 10-fold cross-validation.

Main Results:

• The original model demonstrated lower accuracy for Black, Native Hawaiian/Other Pacific Islander, Asian, and unknown racial/ethnic groups compared to White individuals.
• Fairness approaches improved accuracy for Black individuals but significantly decreased sensitivity in 5 of 6 methods.
• Sensitivity declined by up to 31% for Black individuals in several fairness-improved models.

Conclusions:

• Fairness-improving models showed mixed results, improving some metrics but worsening others, indicating the original model was not truly improved.
• Reduced sensitivity and negative predictive value in fairness-improved models could lead to under-identification of high-risk individuals.
• Implementing these fairness-adjusted models may inadvertently perpetuate racial/ethnic inequities in preventive service allocation.