The association of demographic factors with negative descriptors in medical text
Additional files of interest
Project team: Andrew Zimolzak MD MMSc, Traber Giardina Ph.D. MSW, Darius Dawson Ph.D., Terri Fletcher Ph.D., Taylor Scott, Debra Choi, Ph.D.
Background
Prior authors have critiqued the choice of words in medical texts (PMID 9412299), specifically stigmatizing and biased language. Only recently has electronic text analysis made it possible to characterize clinical language usage on a large scale. Recent work (PMID 35044842) used an expert panel to develop a list of negative patient descriptors and found that Black patients had disproportionately higher odds of negative patient descriptors appearing in the history and physical notes of their medical records compared with White patients.
In VA data, we have previously defined several electronic triggers to detect patterns suggestive of diagnostic error. The trigger catalog includes emergency department (ED) visits with dizziness or abdominal pain with high-risk features, ED return visits (PMID 33180032) (34279630), and cancer-focused triggers for lung and colorectal cancer. Our team has experience refining and running these triggers in the VA data warehouse and translating to other sites' Epic warehouses.
Specific Aims
-
We will select and implement one trigger using available data.
-
We will use a previously tested word list to measure the association of negative patient descriptors with marginalized group membership. We will compare the association in two strata: those with missed opportunities of diagnosis versus those without.
-
We will perform an open-ended text analysis to discover new words associated with membership in a stigmatized group.
Approach
Leveraging the existing ED triggers, we will use a previously developed list of terms, negative patient descriptors (PMID 35044842), to find the base rates of those terms occurring in notes with suspect diagnostic inaccuracies and measure the associations specified above in aim 2. Time permitting, we will randomly sample notes and label them for negative, positive, or unrelated uses to estimate rates of these types of usages. For aim 2, we will also derive a new word list partly from additional work (PMID 34259849) and partly from new hypotheses we derive. For aim 3, we will use log-likelihood (Dunning T, 1993), commonly used for statistical analysis of texts, to determine what words of any type are used in one group versus another, rather than a pre-specified word list. We will analyze racial bias in this initial work. However, other factors for future analysis could include gender, body shape, and mental health comorbidity. Eventually, we could retrain a model to detect positive, negative, or irrelevant word uses. We could also review a sample of notes for themes (PMID 34259849).
Data requested (also specify study population including inclusion/exclusion criteria)
The study population comprises all patients with an emergency department visit and ED discharge to home, followed by hospital admission within 30 days of the index date. If needed, we can limit to a specific date range or diagnoses in collaboration with IT. We request access to demographics (including age, gender, race, ethnicity), problem list ICD codes, encounter data (dates, ICD codes, and specialty), procedures, and clinical notes. We request patient weight, height, and BMI if easy to retrieve.
Expected outcomes/benefits & Organizational Impact
Characterizing the institutional experience can assess local experience with missed opportunities of diagnosis and associated factors. This project serves BCM's commitment to diversity, equity, and inclusion; and the Commitment to Racial Justice to Improve Public Health. In particular, this relates to position statement points 3 (Educate our healthcare workers, scientists, and communities to address structural racism and mitigate its negative impacts on health) and 5 (Address inequitable health outcomes affecting underrepresented minorities).
Do you require support for data analysis once the data is provided (Y/N)? If so, provide details. No.