Toward Automated Detection of Biased Social Signals from the Content of Clinical Conversations

Presentation Time: 09:45 AM - 10:00 AM

Abstract Keywords: Natural Language Processing, Health Equity, Racial disparities, Patient Engagement and Preferences, Deep Learning
Primary Track: Applications

Implicit bias can impede patient-provider interactions and lead to inequities in care. Raising awareness is key to reducing such bias, but its manifestations in the social dynamics of patient-provider communication are difficult to detect. In this study, we used automated speech recognition (ASR) and natural language processing (NLP) to identify social signals in patient-provider interactions. We built an automated pipeline to predict social signals from audio recordings of 782 primary care visits that achieved 90.1% average accuracy across codes, and exhibited fairness in its predictions for white and non-white patients. Applying this pipeline, we identified statistically significant differences in provider communication behavior toward white versus non-white patients. In particular, providers expressed more patient-centered behaviors towards white patients including more warmth, engagement, and attentiveness. Our study underscores the potential of automated tools in identifying subtle communication signals that may be linked with bias and impact healthcare quality and equity.

Speaker(s):
Feng Chen, MS
University of Washington

Author(s):
Feng Chen, MS - University of Washington; Manas Satish Bedmutha, BS; Ray-Yuan Chung, MPH - University of Washington; Janice Sabin - University of Washington; Wanda Pratt, PhD, FACMI - University of Washington; Brain Wood, MD - University of Washington; Nadir Weibel, PhD - UC San Diego; Andrea Hartzler, PhD - University of Washington; Trevor Cohen, MBChB, PhD - Biomedical Informatics and Medical Education, University of Washington;

Using Large Language Models to Detect Stigmatizing Language in Clinical Documentation

Presentation Time: 10:00 AM - 10:15 AM

Abstract Keywords: Fairness and Elimination of Bias, Natural Language Processing, Large Language Models (LLMs)
Working Group: Natural Language Processing Working Group
Primary Track: Applications
Programmatic Theme: Clinical Research Informatics

Stigmatizing language (SL) in electronic health records (EHRs) is widespread and has been measured to capture
implicit biases within clinical notes. The use of SL significantly influences outcomes for individuals experiencing
substance use disorders (SUDs) and various other chronic conditions.1 Previous research on understanding biases was performed using qualitative interviews of patients and healthcare providers.1 Recent advancements in natural language processing (NLP) have made it possible to analyze SL in clinical notes from EHRs.2 The National Institute on Drug Abuse’s (NIDA) “Words Matter”3 initiative provided a list of terms and phrases to avoid during clinical documentation and patient-physician communication. However, it is important to note that while many of these terms may appear in clinical notes, they can sometimes be used in different contexts that do not necessarily constitute SL. Therefore, it is imperative to understand the contextual nuances of these terms to accurately identify and characterize the presence of SL in clinical notes. Large language models (LLM) have gained popularity for their efficacy in understanding languages, marking significant milestones in the field of NLP. These models excel in various tasks including text generation and question answering. We developed an automated method to detect and characterize SL in clinical notes using an advanced fine-tuned LLM called “Fine-tuned LAnguage Net-T5 (FLAN-T5)”.4 We used FLAN-T5 in a “question-answering” fashion to help extract SL from clinical texts.

Speaker(s):
Braja Gopal Patra, PhD
Weill Cornell Medicine; Dept of Population Health Sciences; Div of Health Informatics

Author(s):
Prakash Adekkanattu, PhD - Weill Cornell Medicine; Veer Vekaria; Marianne Sharko, MD, MS - Weill Cornell Medical College Health Policy and Research; Jessica Ancker, MPH, PhD, FACMI - Vanderbilt University Medical Center; Jyotishman Pathak, PhD - Weill Cornell Medical College, Cornell University;

Does Cohort Selection Affect Machine Learning from Clinical Data

Presentation Time: 10:15 AM - 10:30 AM

Abstract Keywords: Machine Learning, Fairness and elimination of bias, Racial disparities, Clinical Decision Support
Primary Track: Applications
Programmatic Theme: Clinical Informatics

This study investigates cohort selection and its effects on the quality of machine learning (ML) models trained on clinical data. It discusses the potential repercussions of making arbitrary decisions during data processing prior to applying ML methods. Experiments are performed within the framework of the National COVID Cohort Collaborative (N3C) dataset. The research aims to unravel biases and assess the fairness of machine learning models used to predict outcomes for hospitalized patients. Detailed discussions follow, covering the data, decision-making processes, and the resulting impact on model predictions regarding patient outcomes. An experiment is conducted in which four arbitrary decisions are made, resulting in 16 distinct datasets characterized by varying sizes and properties. The findings demonstrate significant differences in the obtained datasets and indicate a high potential for bias based on inclusion or exclusion decisions. The results also confirm significant differences in the performance of models constructed on different cohorts, especially when cross-compared between ones based on different inclusion criteria.

Speaker(s):
Atefehsadat Haghighathoseini, Doctoral Student
George Mason University

Author(s):

Adaptive Recruitment Resource Allocation to Improve Cohort Representativeness in Participatory Biomedical Datasets

Presentation Time: 10:30 AM - 10:45 AM

Abstract Keywords: Diversity, Equity, Inclusion, Accessibility, and Health Equity, Fairness and elimination of bias, Administrative Systems, Legal, Ethical, Social and Regulatory Issues, Data Mining
Primary Track: Foundations
Programmatic Theme: Clinical Research Informatics

Large participatory biomedical studies – studies that recruit individuals to join a dataset – are gaining popularity and investment, especially for analysis by modern AI methods. Because they purposively recruit participants, these studies are uniquely able to address a lack of historical representation, an issue that has affected many biomedical datasets. In this work, we define representativeness as the similarity to a target population distribution of a set of attributes and our goal is to mirror the U.S. population across distributions of age, gender, race, and ethnicity. Many participatory studies recruit at several institutions, so we introduce a computational approach to adaptively allocate recruitment resources among sites to improve representativeness. In simulated recruitment of 10,000-participant cohorts from medical centers in the STAR Clinical Research Network, we show that our approach yields a more representative cohort than existing baselines. Thus, we highlight the value of computational modeling in guiding recruitment efforts.

Speaker(s):
Victor Borza
Vanderbilt University Department of Biomedical Informatics

Author(s):
Andrew Estornell, Ph.D. - ByteDance Research; Ellen Wright Clayton, MD, JD - Vanderbilt Medical Center; Chien-Ju Ho, Ph.D. - Washington University in St. Louis; Russell Rothman, M.D., M.P.P. - Vanderbilt University Medical Center; Yevgeniy Vorobeychik, Ph.D. - Washington University in St. Louis; Bradley Malin, PhD - Vanderbilt University Medical Center;

Examining Racial Disparities in Automatic Speech Recognition Performance: Potential Confounding by Provenance

Presentation Time: 10:45 AM - 11:00 AM

Abstract Keywords: Fairness and Elimination of Bias, Human-computer Interaction, Deep Learning
Primary Track: Applications

Automatic speech recognition (ASR) has significantly advanced through the implementation of self-supervised learning and pre-training techniques and can readily adapt to various downstream tasks and applications including health-related tasks. Driven by a growing concern for racial disparities, our experiments demonstrate that ASR performance can be substantially improved by using dialect-specific corpora as one way towards maximizing the utility and reliability of ASR models in digital healthcare workflows. Furthermore, our study indicates that audio recording quality may negatively impact ASR performance and needs to be taken into account as a potential confounding factor in studying racial disparities.

Speaker(s):
Changye Li, PhD
University of Minnesota

Author(s):
Changye Li, PhD - University of Minnesota; Trevor Cohen, MBChB, PhD - Biomedical Informatics and Medical Education, University of Washington; Serguei Pakhomov, PhD - University of Minnesota;