Integration of fairness-awareness into clinical language processing models.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Health Informatics

Background:

• Equitable deployment of clinical AI requires consistent performance across diverse populations.
• Missing/inconsistent race data in EHRs hinders cohort representation and bias assessment.
• This study evaluates AI model performance and fairness in predicting race from clinical text.

Purpose of the Study:

• To compare deep learning models for race prediction from clinical text.
• To assess the impact of fairness-aware optimization on model equity.
• To identify architectural and systemic factors contributing to algorithmic bias.

Main Methods:

• Compared four transformer models and one hierarchical CNN using a two-phase active learning framework.
• Applied a fairness-aware loss function to mitigate racial disparities.
• Evaluated performance and equity via 10-fold cross-validation and subgroup audits.

Main Results:

• Hierarchical CNN achieved higher accuracy and equity (macro F1 = 98.4%) than transformers.
• Fairness constraints improved parity in transformers but degraded hierarchical model performance.
• Persistent disparities indicated architectural limitations and systemic biases.

Conclusions:

• Fairness integration in clinical NLP models is feasible but model-dependent.
• Architectures aligned with clinical text structure inherently promote fairness.
• Upstream documentation inequities drive algorithmic bias, necessitating tailored interventions.