MedDialog-EN Dataset

• MedDialog-EN is a large-scale English dataset featuring over 257K consultations structured into initial condition descriptions and balanced patient-doctor dialogues.
• It captures detailed multi-turn interactions across 51 community categories and 96 specialties, facilitating robust natural language and dialogue modeling in telemedicine.
• Its open access design and simplified text format support applications in virtual diagnostics, triage automation, and remote healthcare consulting.

MedDialog-EN is a large-scale English-language dataset of patient–doctor consultations designed to support machine learning research in medical dialogue systems and telemedicine. Structured as a combination of rich medical condition descriptions and multi-turn conversations across a broad array of specialties, MedDialog-EN has established itself as a foundational resource for real-world conversational AI in healthcare contexts.

1. Dataset Composition and Structure

MedDialog-EN contains 257,454 individual consultations between patients and medical professionals, totaling 514,908 utterances. Each consultation consists of two distinct components:

• An initial textual description of the patient’s medical condition, which may include health history and current symptoms.
• An ensuing dialogue between the patient and the doctor, structured as alternating turns (257,454 patient utterances; 257,454 doctor utterances).

The consultations span a wide variety of clinical contexts:

• 51 community categories, including diabetes, elderly health concerns, and pain management.
• 96 specialties, such as andrology, cardiology, nephrology, and pharmacology.
• The temporal range of data collection spans from 2008 to 2020.

This structural design supports both context-rich natural language understanding and downstream modeling tasks such as response generation and clinical prediction. Each record is designed to facilitate the progressive intake of patient history, symptomatology, and the back-and-forth of virtual consultation.

2. Purpose and Principal Applications

MedDialog-EN was created to advance the development of medical dialogue systems, particularly within the telemedicine domain. The dataset’s primary utility lies in enabling statistical and neural models to:

• Simulate “virtual doctor” capabilities by conducting plausible, informative, multi-turn conversations with patients.
• Improve access to clinical expertise for underserved populations by underpinning remote consultation services.
• Reduce medical system workload by augmenting or automating initial consultation and triage tasks, thereby targeting improvements in healthcare quality, accessibility, and cost efficiency.

A further application lies in decision support for medical professionals—by training dialogue agents that can assist with the management of routine consultations, the dataset contributes to potential reductions in physician burnout and facilitates broader dissemination of clinical guidance.

3. Scope, Specialties, and Coverage

A distinctive feature of MedDialog-EN is its breadth of coverage across specialties and problems:

• 51 community-defined medical categories and 96 specialties offer extensive topical diversity.
• The equal split of over 500,000 utterances ensures balanced representation of both patient concerns and physician responses, making MedDialog-EN suitable for modeling both user (patient) and agent (doctor) personas.
• Compared to its sister dataset, MedDialog-CN (Chinese; 1.1M consultations, 4M utterances, with up to 172 fine-grained specialties), MedDialog-EN focuses on English-speaking populations with cross-national utility.

A summary comparison:

Dataset	No. Consultations	No. Utterances	No. Specialties
MedDialog-EN	257,454	514,908	96
MedDialog-CN	1,145,231	3,959,333*	172**

*Counts for MedDialog-CN may reduce to 3,209,660 if merging consecutive utterances by the same speaker. **172 fine-grained specialties in MedDialog-CN; 96 in MedDialog-EN.

MedDialog-EN’s focus on breadth, diversity, and clinical relevance makes it a valuable benchmark and pretraining corpus for medical conversational AI targeting the anglophone world.

4. Technical Features and Accessibility

MedDialog-EN is characterized by its scale and simplicity of design. The dataset:

• Lacks highly granular semantic or sub-utterance labels, instead providing clean, large-scale, contextually coherent dialogue examples.
• Is distributed through a publicly available GitHub repository: https://github.com/UCSD-AI4H/Medical-Dialogue-System
• Is intended for research use only, with the paper noting that the data is “open to the public”—researchers should verify the repository for any updates or additional licensing stipulations.

The dataset’s plain text format and standard division into description and dialogue sections ease its ingestion for diverse modeling strategies, including supervised learning for response selection, sequence-to-sequence generation, joint embedding models, and few-shot/fine-tuning pipelines.

5. Research Impact and Comparison to Related Resources

MedDialog-EN is cited as the largest medical dialogue dataset available in English at its time of publication, supporting a range of telemedicine applications. Compared to MedDialog-CN (which is larger and more fine-grained in Chinese), MedDialog-EN’s strength lies in its applicability to English-language systems and its facilitation of cross-lingual and comparative studies.

Its distinction from other datasets such as ReMeDi and IMCS-21 is primarily in scale and annotation depth:

• MedDialog-EN: Large-scale, English, broad-specialty, basic structure with strong focus on real-world consultation context.
• ReMeDi: Fewer dialogues but rich sub-utterance semantic annotations, supporting NLU, dialogue policy learning, and NLG in addition to NLG.
• IMCS-21: Multi-level granular annotation in Chinese focused on pediatric domains.

For downstream users, MedDialog-EN provides the raw material for broad-coverage, high-volume dialogue modeling, while other resources can be preferred for sequence labeling or detailed policy learning tasks.

6. Limitations and Future Directions

MedDialog-EN, while extensive, is limited in lack of sub-utterance or fine-grained semantic annotation. The paper notes that the dataset is “continuously growing,” and identifies future directions including:

• Expansion of metadata (e.g., additional demographic attributes) to reduce bias.
• Integration of more complex dialogue phenomena, such as multi-turn context with consecutive same-speaker utterances and richer patient history constructs.
• Cross-lingual and multimodal alignment with datasets such as MedDialog-CN for comprehensive telemedicine research.

A plausible implication is that MedDialog-EN’s evolving scale and structure could underpin transfer learning for more semantically intensive datasets and ultimately facilitate the development of more sophisticated and diagnostically robust medical dialogue agents.

7. Summary

MedDialog-EN is a large, publicly accessible English-language dataset of doctor–patient conversations, encompassing over a quarter million consultations and covering 51 community categories and 96 specialties across more than a decade. By providing structured condition descriptions and multi-turn dialogues, it serves as a foundational resource for the development, training, and evaluation of telemedicine systems and virtual medical agents, particularly for the English-speaking world. Its continued growth and potential for integration with more intricate labeling and multimodal signals underscore its significance within the medical AI research ecosystem (He et al., 2020).