Feed

Equitable deployment of clinical artificial intelligence systems requires consistent performance across diverse patient populations. However, race information in electronic health records is often missing/inconsistently documented, limiting the ability to construct representative cohorts or assess algorithmic bias. This study evaluates model performance and fairness in predicting race from clinical text.

We compared four transformer-based deep learning models with a hierarchical convolutional neural network designed to capture the multilevel structure of clinical narratives. A two-phase active learning framework guided annotation of a primary care database. A fairness-aware loss function was applied to mitigate disparities across racial groups. Each model was trained with and without fairness-aware optimization. Performance and equity were evaluated using 10-fold cross-validation and subgroup audits across race, sex, age, and their intersections.

Here we show that the hierarchical convolutional neural network achieves higher accuracy and performance equity than transformer models (macro F1 = 98.4%). Fairness constraints enhance parity across most transformer architectures, but degrade hierarchical model performance and cause one clinical model to collapse toward majority predictions, demonstrating that fairness interventions are highly model dependent. Persistent disparities across race, sex, and age indicate that inequities reflect architectural limitations and systemic biases.

This study demonstrates that fairness can be integrated into clinical language models, though effects vary by model type. Architectures aligned with clinical text structure inherently promote fairness, yet mixed fairness constraint outcomes highlight the need for tailored interventions. Persistent demographic disparities show that algorithmic bias often reflects upstream documentation inequities. This framework offers a scalable path toward equitable NLP for clinical artificial intelligence.