![[ Visit Client Website ]](images/banner.gif)
Schedule:
Friday, January 22, 2027
Continental 1-4 (Hilton San Francisco Union Square)
* noted as presenting author
Background and Purpose:
Hospitalization is a critical opportunity for interdisciplinary teams, including social workers, to document, discuss, and coordinate harm-reduction (HR) care for patients with opioid use disorder (OUD). Clinical documentation can shape care planning, referrals, and discharge preparation, yet the language used in notes may not be distributed equitably across racial and ethnic groups. Because documentation language offers a measurable proxy for whether HR concepts were addressed during hospitalization, natural language processing (NLP) can help identify inequities that may otherwise remain invisible. This study used NLP to examine whether HR-related clinical language in hospital notes for patients with OUD differed across racial and ethnic groups.
Methods:
We analyzed 9,965 clinical notes from 2,182 hospitalized patients with OUD diagnoses (ICD-10: F11.x) in the MIMIC-III/IV datasets from Beth Israel Deaconess Medical Center, spanning 2008–2019. A 631-term HR vocabulary covering six domains, including MOUD management, overdose prevention, risk-reduction education, HR referrals, other pharmacological supports, and psychosocial or environmental supports, was developed through expert-informed consensus and corpus-based semantic expansion, with candidate terms reviewed for contextual relevance and domain accuracy. Race and ethnicity were recoded from 33 registration values into five analytic categories following OMB standards: Non-Hispanic White (reference), Non-Hispanic Black, Hispanic/Latinx, Asian, and Other/Multiracial. Notes from patients with unknown or declined race and ethnicity were excluded. The primary outcome was the binary presence of any HR vocabulary in a note; secondary outcomes were category-specific presence. Multivariable logistic regression with cluster-robust standard errors estimated adjusted associations between race and ethnicity and HR language, controlling for age, gender, and note type (analytic sample: 7,307 notes from 1,854 patients). As a robustness check, we also estimated patient-level models aggregating note content within patients. Sensitivity analyses therefore assessed whether observed disparities reflected repeated note structure rather than a single documentation pattern within individuals alone.
Results:
HR vocabulary appeared in 57.6% of notes overall. Pharmacological language was most common (33.6%), while psychosocial or environmental language appeared in only 2.1% of notes. In adjusted models, significant racial and ethnic disparities were concentrated in pharmacological HR documentation. Compared with Non-Hispanic White patients, Non-Hispanic Black patients had 36.5% lower odds of pharmacological HR vocabulary overall (OR=0.635, 95% CI: 0.492–0.820, p<0.001), including lower odds of MOUD management language (OR=0.786, p=0.031), overdose prevention language (OR=0.696, p=0.002), and other pharmacological language (OR=0.761, p=0.013). Hispanic/Latinx patients had 39.9% lower odds of overdose prevention vocabulary (OR=0.601, 95% CI: 0.422–0.856, p=0.005). Findings were substantively similar in patient-level sensitivity analyses.
Hospitalization is a critical opportunity for interdisciplinary teams, including social workers, to document, discuss, and coordinate harm-reduction (HR) care for patients with opioid use disorder (OUD). Clinical documentation can shape care planning, referrals, and discharge preparation, yet the language used in notes may not be distributed equitably across racial and ethnic groups. Because documentation language offers a measurable proxy for whether HR concepts were addressed during hospitalization, natural language processing (NLP) can help identify inequities that may otherwise remain invisible. This study used NLP to examine whether HR-related clinical language in hospital notes for patients with OUD differed across racial and ethnic groups.
Methods:
We analyzed 9,965 clinical notes from 2,182 hospitalized patients with OUD diagnoses (ICD-10: F11.x) in the MIMIC-III/IV datasets from Beth Israel Deaconess Medical Center, spanning 2008–2019. A 631-term HR vocabulary covering six domains, including MOUD management, overdose prevention, risk-reduction education, HR referrals, other pharmacological supports, and psychosocial or environmental supports, was developed through expert-informed consensus and corpus-based semantic expansion, with candidate terms reviewed for contextual relevance and domain accuracy. Race and ethnicity were recoded from 33 registration values into five analytic categories following OMB standards: Non-Hispanic White (reference), Non-Hispanic Black, Hispanic/Latinx, Asian, and Other/Multiracial. Notes from patients with unknown or declined race and ethnicity were excluded. The primary outcome was the binary presence of any HR vocabulary in a note; secondary outcomes were category-specific presence. Multivariable logistic regression with cluster-robust standard errors estimated adjusted associations between race and ethnicity and HR language, controlling for age, gender, and note type (analytic sample: 7,307 notes from 1,854 patients). As a robustness check, we also estimated patient-level models aggregating note content within patients. Sensitivity analyses therefore assessed whether observed disparities reflected repeated note structure rather than a single documentation pattern within individuals alone.
Results:
HR vocabulary appeared in 57.6% of notes overall. Pharmacological language was most common (33.6%), while psychosocial or environmental language appeared in only 2.1% of notes. In adjusted models, significant racial and ethnic disparities were concentrated in pharmacological HR documentation. Compared with Non-Hispanic White patients, Non-Hispanic Black patients had 36.5% lower odds of pharmacological HR vocabulary overall (OR=0.635, 95% CI: 0.492–0.820, p<0.001), including lower odds of MOUD management language (OR=0.786, p=0.031), overdose prevention language (OR=0.696, p=0.002), and other pharmacological language (OR=0.761, p=0.013). Hispanic/Latinx patients had 39.9% lower odds of overdose prevention vocabulary (OR=0.601, 95% CI: 0.422–0.856, p=0.005). Findings were substantively similar in patient-level sensitivity analyses.
Conclusions and Implications:
Racial and ethnic disparities in HR-related clinical documentation are detectable in routine hospital notes. Although note language is not a direct measure of services delivered, it is a meaningful and scalable indicator of whether overdose prevention, MOUD, and related HR concepts were documented in care planning. The especially pronounced gap in overdose prevention documentation identifies an actionable target for equity-focused hospital practice. Social work researchers and practitioners can use NLP-based equity audits, structured documentation prompts, and discharge planning protocols to identify and reduce racialized gaps in HR communication and coordination.