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. Yet, 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 systematically investigate persistent knowledge gaps in mismatched data coding themes and specialized languages, and also highlight how clinical data can act as boundary objects and human knowledge brokers to alleviate these challenges. Our work underscores the benefits of leveraging analytical frameworks like AT to systematically examine interdisciplinary teams' collaborative work and provide meaningful insights on best practices in future collaboration.
Introduction
The integration of Artificial Intelligence (AI) into healthcare promises to revolutionize patient care, but its success hinges on the "seamless" partnership between clinicians and technical researchers. Despite the potential, these interdisciplinary teams frequently encounter friction that leads to project failure. While individual case studies have documented these challenges, the research community has lacked a systematic framework to analyze how these barriers manifest across complex workflows. Clinicians naturally prioritize patient care and nuanced symptom observation, while technical experts focus on model performance and structured data schemas. Without a shared analytical lens, these divergent goals lead to mismatched data and ineffective tools.
Methodology
To investigate these dynamics, we conducted a qualitative study involving semi-structured interviews with 13 experts involved in two telehealth-based AI projects within the domain of speech-language pathology (SLP). The participant pool consisted of six clinicians (Speech-Language Pathologists) and seven AI/NLP experts, ranging from junior researchers to senior practitioners.
The primary data for this analysis included transcribed interviews and the clinical artifacts used during collaboration, such as audio/video recordings of patient interactions, transcripts, and manual clinical coding spreadsheets.
The Activity Theory Framework
We utilized Activity Theory (AT), specifically Engeström’s "Activity Triangle," as our primary analytical lens. This framework allows for the deconstruction of collaborative work into seven interconnected elements: Subject, Object, Tools, Rules, Community, Division of Labor, and Outcome.
By mapping the collaboration onto this triangle, we identified specific "tension points" where clinical and technical workflows intersected but failed to align. For example, institutional "Rules" (such as HIPAA compliance) often created friction with the "Tools" and "Division of Labor" required for flexible AI data labeling.
Key Findings
Data as a Double-Edged Sword
Clinical data serves as the central "Object" shared by both groups, yet it undergoes divergent transformations. Clinicians code data to plan patient interventions, while technical experts transform that same data into model inputs. We found that if the clinical taxonomy does not map directly to the algorithm's schema, the collaboration often stalls due to a lack of shared understanding of what the data represents.
The Power of Boundary Objects and Knowledge Brokers
To bridge these gaps, successful teams utilize boundary objects—concrete artifacts like specific spreadsheets or video segments. These objects are "plastic" enough to be interpreted differently by each group but "robust" enough to maintain a shared identity, serving as a reliable alternative to jargon-heavy verbal explanations.
Furthermore, we identified the critical role of the "Bilingual" Knowledge Broker. These individuals possess three essential skills:
- Familiarity with the knowledge boundaries of both domains.
- The ability to identify miscommunications in real-time.
- The skill to translate specialized jargon into accessible language for all stakeholders.
Conclusion and Future Directions
This study provides a qualitative roadmap for interdisciplinary clinical AI work. By visualizing the network of actions and identifying the characteristics of effective brokers, we move beyond vague calls for "better communication" toward structural solutions.
Future work should explore the potential for AI-powered dynamic boundary objects, where Large Language Models (LLMs) might act as automated brokers, translating clinical nuances into technical requirements in real-time. Additionally, longitudinal studies are needed to track how these collaborative structures evolve over the multi-year lifecycle of large-scale industrial AI projects.
Summary updated