How to find and reduce demographic bias in an NLP model trained on free-text clinical notes, using pediatric anxiety detection as the case study. Learners see how outcome parity is measured across demographic subgroups (accuracy and false-negative-rate gaps), how the source of bias is located by analysing differences in word distributions and information density between groups rather than in structured features, and how a data-centric de-biasing intervention - neutralizing demographically-associated terms while preserving clinically salient content - lowers the diagnostic disparity without discarding signal. Frames the four bias types (selection, label, textual, over-amplification) and why unstructured healthcare text needs tailored fairness methods.
Detects and mitigates demographic (gender) bias in the unstructured clinical-note text used to train an AI model for pediatric anxiety screening. The authors measure outcome parity across sex subgrou…