Intersectional and Marginal Debiasing in Prediction Models for Emergency Admissions.

Area of Science:

• Health Informatics
• Clinical Decision Support
• Health Equity Research

Background:

• Fair clinical prediction models are essential for equitable health outcomes.
• Existing fair algorithms often use marginal debiasing, simplifying patient subgroups.
• This simplification may not adequately address discrimination in intersectional patient groups.

Purpose of the Study:

• To evaluate the impact of simplifying patient subgroups during training on intersectional subgroup performance in emergency department (ED) admission prediction models.
• To compare the effectiveness of intersectional debiasing versus marginal debiasing strategies.

Main Methods:

• A prognostic study utilizing retrospective data from two large ED cohorts (MIMIC-IV and BCH).
• Admission prediction models were trained using variations in fairness optimization (marginal vs. intersectional debiasing).
• Performance was assessed using metrics like area under the receiver operator characteristic curve (AUROC), calibration error, and false-negative rates across subgroups defined by race, ethnicity, and gender.

Main Results:

• Intersectional debiasing significantly reduced subgroup calibration error and false-negative rates compared to marginal debiasing in both cohorts.
• For example, in the MIMIC-IV cohort, intersectional debiasing reduced calibration error by 22.3% compared to 11.3% with marginal debiasing.
• These fairness improvements did not compromise overall model accuracy, with AUROC remaining consistent with baseline models.

Conclusions:

• Intersectional debiasing is more effective than marginal debiasing in mitigating performance disparities across intersecting patient groups for ED admission prediction.
• Models developed with intersectional debiasing achieved reduced group-specific errors without sacrificing overall predictive accuracy.
• Incorporating intersectional debiasing into the development of clinical risk prediction models is recommended to promote health equity.