Abstract
Medical conversation is a central part of medical care. Yet, the current state and quality of medical conversation is far from perfect. Therefore, a substantial amount of research has been done to obtain a better understanding of medical conversation and to address its practical challenges and dilemmas. In line with this stream of research, we have developed a multi-layer structure annotation scheme to analyze medical conversation, and are using the scheme to construct a corpus of naturally occurring medical conversation in Chinese pediatric primary care setting. Some of the preliminary findings are reported regarding 1) how a medical conversation starts, 2) where communication problems tend to occur, and 3) how physicians close a conversation. Challenges and opportunities for research on medical conversation with NLP techniques will be discussed.
Introduction
Medical consultations are the heartbeat of healthcare, yet they are notoriously difficult to analyze at scale. When a doctor says "Okay," are they agreeing with a patient’s concern, or simply signaling that they are listening before moving to the next topic? To build NLP systems that can truly assist in the exam room, we need to understand the "deep structure" of these conversations.
Most existing medical NLP research treats dialogue as a flat sequence of isolated utterances. However, human conversation is inherently interactive and non-linear. A patient might resist a treatment recommendation, leading to a "negotiation" that spans several turns. Researchers Nan Wang, Yan Song, and Fei Xia identify two major gaps: resource scarcity for natural medical dialogue and a neglect of structural interactivity in current coding frameworks. Without this context, NLP models struggle with tasks like automated summarization or extracting information for Electronic Health Records (EHRs).
The COSTA Framework
The core contribution of this work is the COSTA (COnversational STructures and Actions) scheme. It treats conversation like a tree structure, similar to how a dependency parser treats a sentence. The hierarchy consists of:
- Phases: High-level stages like Opening, History-taking, and Treatment Recommendation.
- Sequences: Clusters of related turns (e.g., a "News Delivery Sequence").
- Pairs: "Adjacency pairs" where one turn triggers another (e.g., Question $\rightarrow$ Answer).
- Turns: The individual units of speech, including para-linguistic features like silence, overlaps, and nodding.
By labeling these layers, the researchers can track how a conversation "transits" from one phase to another and identify which phases are the most complex or prone to friction. As shown in Figure 1, sequences can be expanded through insert or pre-expansion pairs, reflecting the dynamic nature of the dialogue.
The Chinese Pediatric Corpus
The study introduces a significant new resource: a corpus of 1,000 video-recorded and manually transcribed medical conversations from Chinese pediatric primary care settings. The subset analyzed in this paper includes 187 visits involving 9 physicians, totaling over 275,000 characters and 23,000 turns. Unlike many Western datasets, this focuses on Chinese clinical settings, providing a rare look at non-Western medical communication norms.
Structural Analysis and Findings
The preliminary analysis of the 187-visit subset yielded concrete structural data regarding the complexity and flow of medical encounters.
The "Negotiation Zone"
The researchers found that the Treatment Recommendation phase (P5) is the most complex, averaging 6.63 turns per sequence. This is where "problems" occur; patients often withhold immediate acceptance of a diagnosis or prescription, forcing doctors to expand the sequence to secure agreement. In contrast, the Opening phase (P0) is the simplest, with only 2.15 turns per sequence.
Cultural Openings and Closings
In a striking contrast to U.S. data (where social openings occur in less than 10% of cases), 33% of Chinese pediatric visits began with a social "Opening Phase" (greetings, identity confirmation). This suggests that rapport-building is a more explicit structural requirement in this medical culture. However, 67% of visits still skip social pleasantries and go straight to the "Problem Presentation."
Regarding closings, doctors initiate the end of a visit primarily by making future arrangements (52%) or summarizing treatment plans (38%). Only 10% of closings were "abrupt" announcements that the visit was over.
Conclusion and Future Directions
The COSTA framework demonstrates that context is king. A word like "Yeah" has entirely different meanings depending on its position in the COSTA tree. If it's the second part of a "Treatment-Acceptance" pair, it’s a decision; if it’s in the middle of a doctor’s long explanation, it’s just a "continuer."
While manual annotation is labor-intensive, the authors plan to use this "Treebank" of conversations to train automatic conversation parsers. Future work will also involve mapping specific conversation phases directly to sections of the Electronic Health Record (e.g., mapping the "Problem Presentation" phase to the "Symptom" section of a clinical note) and releasing the annotated corpus to the broader research community.
Summary updated