Diversity, Equity and Inclusion (DEI)
Allison N. Wallingford, MD, MSE
Assistant Professor of Clinical Practice
University of Washington
Seattle, Washington, United States
Amy J. Starosta, PhD
Associate Professor
University of Washington
Seattle, Washington, United States
Cherry Junn, MD
Associate Professor
University of Washington
Seattle, Washington, United States
Andrew Humbert, PhD
Assistant Professor
University of Washington
Seattle, Washington, United States
Allison N. Wallingford, MD, MSE
Assistant Professor of Clinical Practice
University of Washington
Seattle, Washington, United States
Quantify biased language describing patients with Traumatic Brain Injury (TBI), especially those from systemically disadvantaged groups, with the goal of increasing provider awareness, reducing biased language, and improving outcomes.
Design:
Multiple methods were used to quantify and compare biased language in the charts of adult patients admitted to our trauma center after TBI and a propensity matched cohort of patients admitted with other trauma. A custom algorithm counted the frequency of five biased words (refuse, addict, combative, uncooperative, and non-compliant) in patient charts. We also used an Artificial Intelligence analysis tool, “VADER”, to score sentiment in chart excerpts. As a preliminary test, VADER scored vignettes that were deliberately written in biased or neutral language. Finally, two trainees are manually reviewing a smaller sample of notes and scoring sentiment.
Results:
The average frequency of each biased word and total biased words did not differ between the two cohorts (p=0.65). The average frequency of any of the five words was 0.8 ± 3.7 words/day of admission for patients with TBI and 0.8 ± 4.3 words/day for patients without TBI. Further analysis with logistic regression of biased word counts will examine whether TBI diagnosis is a factor in biased word occurrence when accounting for other demographics. VADER scores aligned poorly with provider perception of sentiment; neutrally written vignettes were scored as more negative than biased vignettes.
Conclusions:
The selected biased words occurred quite frequently on average in patient charts, although with high variability. However, there was no difference in overall frequency for patients with TBI and patients without TBI. VADER had poor ability to detect language that was considered negative in medical records based on our initial testing. Once manual review of charts is completed, we will compare sentiment scores for patients with and without TBI and determine correlation between sentiment scores and biased word frequency.
