Exploring Interdisciplinary Team Collaboration in Clinical NLP Projects Through the Lens of Activity Theory

Bingsheng Yao, Yao Du, Yue Fu, Xuhai Xu, Yanjun Gao, Hong Yu, Dakuo Wang

Abstract

Interdisciplinary clinical AI projects often face significant challenges due to differing professional languages and priorities between clinicians and engineers. This paper applies Activity Theory (AT) to systematically map these friction points within Clinical Natural Language Processing (NLP) projects. Through a qualitative study of 13 participants, the research identifies how "boundary objects" and "knowledge brokers" facilitate collaboration. The findings suggest that successful interdisciplinary work requires bridging divergent data taxonomies and navigating structural tensions such as institutional rules and divisions of labor.

Introduction

The "Interdisciplinary Gap" is a persistent hurdle in clinical AI development. While clinicians prioritize patient care, medical nuance, and regulatory compliance, AI experts focus on model architecture, performance metrics, and data schemas. This misalignment often leads to a "mismatch of labor and responsibility," where fundamental questions—such as data cleaning responsibilities or the mapping of clinical taxonomies to machine learning labels—remain unresolved. This paper provides a systematic analysis of these failures and the coping mechanisms employed by successful teams.

Theoretical Framework: Activity Theory

To analyze these complex interactions, the researchers utilized Activity Theory (AT). AT provides a conceptual lens that views human activity as a system mediated by various components.

The Activity Theory triangle framework illustrating the relationship between subjects, objects, tools, rules, community, and division of labor.

As shown in the framework, an activity is driven by a Subject (the worker) toward an Object (the goal), mediated by Tools (artifacts). This process is situated within a Community and governed by Rules and a Division of Labor. Tensions or "contradictions" within this system highlight areas where workflows clash or miscommunications occur.

Methodology

The study employed a qualitative case-study approach, conducting semi-structured interviews (35–60 minutes) with 13 participants involved in telehealth-focused AI collaborations.

Demographic and professional profiles of the 13 study participants.

The participant pool included six clinicians (Speech-Language Pathologists) and seven AI experts, with varying levels of experience. The analysis involved inductive coding of interview transcripts, which were then mapped onto Activity Triangles to visualize where workflows overlapped and where tensions emerged.

Key Insights

Data as a "Double-Edged Sword"

Clinical data serves as the primary "Object" for both groups but undergoes "divergent transformations." Clinicians transform raw observations into diagnostic insights (clinical taxonomy), while AI experts transform the same data into model inputs (technical taxonomy). Friction arises when these taxonomies do not align, leading to confusion over data interpretation.

The Role of Boundary Objects and Knowledge Brokers

Concrete artifacts, such as specific spreadsheet rows or video clips, function as boundary objects. These tools are flexible enough to be interpreted differently by each group yet robust enough to maintain a shared identity, helping to resolve deadlocks when verbal jargon fails.

Furthermore, successful teams rely on knowledge brokers—individuals with grounding in both domains. These "bilingual" experts identify miscommunications early and translate specialized jargon into a shared common language.

Structural Tensions

Activity Theory reveals that friction is often rooted in structural elements:

  • Rules: HIPAA constraints and institutional regulations can limit data sharing and model development.
  • Division of Labor: Misunderstandings often occur regarding task ownership, such as when an AI expert expects "computationally clean" labels that are clinically impossible to produce.

Conclusion and Future Work

While this study is limited by its sample size and specific focus on Speech-Language Pathology, it offers a robust framework for understanding interdisciplinary AI collaboration. Future research may explore the role of Large Language Models (LLMs) as "AI-powered dynamic boundary objects" that can proactively identify and bridge terminological gaps in real-time meetings.

Summary updated

Abstract

Advanced AI technologies are increasingly integrated into clinical domains to advance patient care. The design and development of clinical AI technologies necessitate seamless collaboration between clinical and technical experts. However, such interdisciplinary teams are often unsuccessful, with a lack of systematic analysis of collaboration barriers and coping strategies. This work examines two clinical AI collaborations in the context of speech-language pathology via semi-structured interviews with six clinical and seven technical experts. Using Activity Theory (AT) as our analytical lens, we examine persistent knowledge gaps and collaboration tensions across clinical and technical workflows, and show how clinical data can function as boundary objects while interdisciplinary collaborators may act as knowledge brokers to help address these challenges. Our findings contribute to CSCW research on interdisciplinary teams' data work by showing how shared clinical data, boundary objects, and broker roles shape coordination in early-stage clinical AI collaboration, and by providing insights into best practices for future collaboration.
BibTeX
@misc{yao2024exploring,
  author = {Bingsheng Yao and Yao Du and Yue Fu and Xuhai Xu and Yanjun Gao and Hong Yu and Dakuo Wang},
  title = {{Exploring Interdisciplinary Team Collaboration in Clinical NLP Projects Through the Lens of Activity Theory}},
  year = {2024},
  eprint = {2410.00174},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2410.00174},
}
Figure