Forty-seven percent of pregnancies have stigmatizing language in their clinical notes in an electronic health record cohort

Neha Simha, Hannah Takasuka, Li-Ching Chen, et al.

Summary: This preprint reports a large-scale NLP analysis of 640,345 obstetric clinical notes from 26,178 pregnancies at a single academic medical centre, using a keyword-guided BERT classifier to detect stigmatising language. Stigmatising language was found in 47% of pregnancies, with significantly higher odds among Black patients compared to Asian and White patients, among those with lower educational attainment, and among pregnancies complicated by preterm birth. The study demonstrates that automated NLP methods can systematically quantify documentation bias at scale.

Why it matters: Stigmatising language in clinical notes is not merely a linguistic concern — it can influence downstream clinical decisions, shape trainee attitudes, and perpetuate structural inequities in care. The finding that Black patients and those with lower educational attainment are disproportionately exposed to stigmatising documentation is consistent with well-documented racial disparities in obstetric outcomes. For clinical informatics and health systems, this study illustrates a practical, scalable approach to auditing documentation quality and equity across large EHR datasets — a use case with clear operational value extending well beyond obstetrics.

Expert perspective: This is a methodologically interesting and socially important contribution, though several limitations warrant careful attention before conclusions are treated as definitive.

NLP classifier validity is the central methodological concern. A keyword-guided BERT classifier can achieve reasonable performance, but the abstract does not report precision, recall, F1 scores, or inter-rater agreement against a gold-standard human annotation of a held-out sample. Without this, the reported 47% prevalence figure carries substantial uncertainty. If the classifier has meaningful false-positive rates — plausible when "stigmatising" language overlaps with legitimate clinical descriptors — the prevalence estimate and the between-group comparisons may be systematically distorted. Readers should treat the quantitative estimates as approximate pending external validation.

Single-site academic centre data limits generalisability. Academic medical centres have distinctive patient populations, documentation cultures, and trainee composition. Prevalence and disparity patterns at community hospitals, safety-net institutions, or non-academic settings may differ substantially.

Confounding is a real concern in the disparity findings. Black patients in this cohort may have higher rates of clinical complexity, social risk documentation, or both, which could independently increase exposure to certain language patterns. The association with preterm birth is also worth scrutinising: clinically complex pregnancies generate more notes and more opportunities for any language — stigmatising or otherwise — to appear. Volume of documentation is a confounder that is easy to overlook. Causality cannot be inferred from this cross-sectional, observational design.

Despite these caveats, the study makes a genuine contribution. The scale of the analysis and the application of transformer-based NLP to equity monitoring represent a meaningful advance over prior work relying on manual chart review of small samples. The practical implication — that health systems can and should audit their documentation for bias as a quality-improvement activity — is sound. This is promising work, but preprint status and unvalidated classifier performance mean specific numbers should not be cited as established fact without peer-reviewed validation.

Relevance: Clinical Informatics and Health Data Methods

Source: https://pubmed.ncbi.nlm.nih.gov/42619892/

---