Using Natural Language Processing to Detect Bias in Emergency Department Clinical Notes
PI: Isha Agarwal, MD, ScD, Attending Physician and Faculty Scientist, Department of Emergency Medicine, Maine Medical Center
Co-Is: Elizabeth Scharnetzki, PhD, Faculty Scientist I, MaineHealth Institute for Research, Center for Interdisciplinary Population and Health Research; Tania Strout, PhD, Vice Chair for Research, Department of Emergency Medicine, Maine Medical Center; Annika Schoene, PhD, Research Scientist, Institute for Experiential AI, Northeastern University; Jessica DiBiase, MPH, Research Data Analyst II, MaineHealth Institute for Research, Center for Interdisciplinary Population Health Research
Can you briefly describe your project and what motivated you to engage with the translational science aspect of this work?
Our project investigated how stigma and bias may manifest in the emergency department triage process, with the goal of using natural language processing (NLP) to detect subtle linguistic signals of belonging and exclusion. To do this, we analyzed free-text triage notes to identify patterns that may reflect provider bias.
Clinically, we’ve observed disparities in patient acuity scores that may be influenced by underlying bias. Motivated by these observations, we formed an interdisciplinary team and set out to develop a broadly applicable tool for identifying stigmatizing language across clinical settings. The S-GATS mechanism’s emphasis on disease-agnostic translational science resonated strongly with our vision of creating a generalizable framework to improve equity in healthcare documentation.
How did your interdisciplinary team come together, and what role did collaboration play in advancing your project?
Our team was a blend of pre-existing and newly formed collaborations, including clinicians, social scientists, and computer scientists. The diversity of expertise provided clinical relevance, theoretical grounding, and technical capability, which was essential. We found early-career networks and institutional support vital in bridging disciplines. This team science approach exemplifies translational principles by fostering innovation that’s broadly applicable beyond a single disease or discipline.
In addition to assembling a multidisciplinary team, your project engaged a Stakeholder Advisory Group. How did that come about, and how did it shape your understanding of translational science?
The Stakeholder Advisory Group wasn’t part of our original plan. It was added after an early consultation helped us recognize the importance of integrating stakeholder input and focusing more on intervention development — elements that significantly strengthened our project.
We invited volunteers after presenting our initial work, and the response was enthusiastic. While scheduling was challenging, we addressed this by forming smaller subgroups to maintain ongoing engagement. The advisory group provided invaluable feedback, helping us simplify and clarify our annotation guidelines and develop patient-friendly materials like glossaries.
This process ensured our tools were accessible and meaningful beyond academic audiences, reinforcing that translational science is not just about moving from bench to bedside but about making research broadly usable and responsive to real-world needs. Engaging with stakeholders fundamentally expanded our understanding, grounding our work in the perspectives of those it’s ultimately meant to serve.
What were some challenges you faced and/or accomplishments you experienced during your project?
After overcoming early hurdles — including IRB approval, navigating institutional data sharing policies, and establishing HIPAA-compliant data storage — we built real momentum. Our multi-institutional, collaborative team has met regularly for over a year. Together, we initiated annotation of thousands of clinical notes and began developing a machine learning model to support our research goals. Collaborating with a passionate, all-female, early-career research team has been especially meaningful, bringing energy and shared commitment to the project. Although the short timeline and administrative logistics sometimes made it feel like a race against the clock, this effort laid important groundwork for future NLP initiatives at our institutions.
What advice would you give investigators considering whether or not their project fits within a translational science program like S-GATS?
My best advice is to reach out early and actively engage in conversations with the program teams. Even if you’re unsure whether your project fits neatly within the traditional scope, these discussions can help clarify how your idea aligns and reveal opportunities to adapt your approach that can significantly strengthen your proposal. Being flexible and open to iterative feedback throughout the process is crucial, as projects often evolve when new insights or guidance emerge. It’s also important to embrace a broader understanding of translational science beyond the conventional bench-to-bedside pathway.
Many projects that are disease-agnostic, cross-disciplinary, and aimed at creating generalizable tools or solutions often align well with translational science principles. Being open to this broader perspective can help investigators see the potential fit and maximize the impact of their work.