Clinical NLP · Ahmad0067
MedSynth
Synthetic medical dialogue and SOAP note pairs to train clinical summarisation models without patient privacy risks.
What MedSynth is
MedSynth is a synthetic dataset designed for developing and evaluating clinical dialogue-to-note summarisation models, created by Ahmad Rezaie Mianroodi, Amirali Rezaie, Niko Grisel Todorov, Cyril Rakovski, and Frank Rudzicz. It matches simulated doctor–patient conversations with structured medical documentation in the SOAP (Subjective, Objective, Assessment, and Plan) format across a wide range of diagnoses.
Healthcare providers spend between 52 and 102 minutes each day documenting patient consultations in electronic health records, contributing substantially to clinical burnout. While automated language processing tools can help relieve this administrative burden, machine learning research is often constrained by privacy restrictions on genuine patient records and the narrow clinical scope of existing open datasets.
To address these constraints, the dataset provides data augmentation through more than 10,000 synthetic pairs. This setup allows researchers to train and evaluate natural language processing systems across thousands of conditions without handling personal health identifiers.
How it was made
Generation was guided by diagnostic distributions observed in the IQVIA PharMetrics Plus database, a collection of United States health insurance claims filtered for ICD-10 diagnostic codes. Using these disease distributions alongside samples from the ACI-Bench dataset, the text pairs were produced using a collaborative multi-agent architecture composed of large language models.
The dataset spans 2,000 distinct ICD-10 codes, with five synthetic records generated per code. Each record contains four attributes: the ICD-10 diagnostic code, a textual description of the code, a doctor–patient conversation, and a clinical note structured into standard SOAP headings.
Key facts
- Total instances
- More than 10,000 pairs
- Diagnostic coverage
- 2,000 ICD-10 codes (5 pairs per code)
- Clinical note format
- SOAP (Subjective, Objective, Assessment, and Plan)
- File format
- CSV
- Language
- English
- Foundational sources
- IQVIA PharMetrics Plus and ACI-Bench
What you can use it for
- Medical dialogue summarisation
- Training and assessing natural language processing models that convert doctor–patient conversations into structured clinical notes.
- Note-to-dialogue reconstruction
- Developing and testing models that generate plausible clinical conversations directly from structured SOAP notes.
- Clinical data augmentation
- Augmenting existing clinical natural language processing training sets across 2,000 diagnostic categories without relying on restricted patient data.
Explore the data
No licence declared
The authors have not published a licence, which means no permission to redistribute the data has been granted to anyone. The rows are not served here; read them on the original dataset page.How the columns are distributed
Over 10,240 rows of default/train
Note
string_text- min
- 1,211
- median
- 2,978
- mean
- 2994.76
- max
- 5,048
Dialogue
string_text- min
- 1,897
- median
- 4,502
- mean
- 4536.84
- max
- 17,453
0.02% empty
ICD10
string_text- min
- 3
- median
- 5
- mean
- 4.80
- max
- 7
ICD10_desc
string_text- min
- 5
- median
- 39
- mean
- 43.07
- max
- 159
Strengths and limits
Good at
- Provides broad diagnostic diversity by including five pairs for each of 2,000 distinct ICD-10 codes.
- Carries no risk of exposing confidential patient information because the conversations and notes are entirely synthetic.
- Follows the widely recognised SOAP format used extensively across primary care documentation.
- Shapes data generation around real-world diagnostic frequencies derived from large-scale insurance claims.
Watch out for
- Does not guarantee medical factual accuracy and requires expert evaluation before clinical adoption.
- Cannot fully represent the conversational complexity and unpredictable nature of real-world patient interactions.
- Is restricted to the SOAP structure, omitting other clinical documentation formats used in specialised healthcare settings.
- Reflects the underlying biases present in the large language models used during generation.
- Must not be used directly as a source of verified clinical or medical information.
Questions
Can I look inside MedSynth without downloading it?
Not here: it declares no licence, so no permission to redistribute its rows has been granted. Its description and statistics are still shown.
What licence is MedSynth under?
Its authors have not declared a licence, which means no permission to reuse or redistribute it has been granted.
Does this dataset contain real patient information?
No. The dataset is entirely synthetic and does not contain any personal, private, or sensitive patient records.
Can the dataset be used as a reference for clinical treatment?
No. It is intended solely for training and evaluating machine learning models and must not be used as a reliable source of medical information.
What structure do the generated clinical notes follow?
All clinical notes adhere to the SOAP format, which divides clinical documentation into Subjective, Objective, Assessment, and Plan components.
What information is included in each data entry?
Each entry contains four fields: the ICD-10 code, the ICD-10 code description, the simulated dialogue between clinician and patient, and the resulting SOAP note.
How to cite it
Ahmad Rezaie Mianroodi, Amirali Rezaie, Niko Grisel Todorov, Cyril Rakovski, & Frank Rudzicz. (2025). MedSynth: Realistic, Synthetic Medical Dialogue-Note Pairs. arXiv:2508.01401.
Written from MedSynth's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.