IMD (Italian Medical Dataset) is a cutting-edge resource for Natural Language Processing (NLP) in the medical domain, specifically designed to improve the accuracy and reliability of Question Answering (QA) models in the Italian language. The dataset is split into two primary components:
- IMD-QA: Questions and answers extracted from Italian medical forums, reflecting informal language used by patients and healthcare professionals.
- IMD-MCQA: Multiple-choice questions from Italian medical specialization exams, ideal for training models focused on structured and formal medical queries.
To use the dataset, you can access it via Google Drive and download the required files or Download Now from GitHub. The dataset is provided in JSON format, structured to facilitate integration with Natural Language Processing (NLP) models, including those designed for Question Answering (QA) and Multiple-Choice QA (MCQA) tasks.
📂 Dataset Access:
- IMD-QA (Italian Medical QA Dataset): 🔗 Google Drive Link
- IMD-MCQA (Italian Medical MCQA Dataset): 🔗 Google Drive Link
The composition of IMD is:
| Statistic | IMD-QA | IMD-MCQA |
|---|---|---|
| # Questions and Answers | 782,644 | 25,862 |
| # Categories | 77 | 60 |
| Last Update | July 2024 | July 2024 |
| Total Answer Tokens | 40,370,381 | 9,321 |
| Unique Answer Vocabulary | 154,837 | 1,234 |
| Total Question Tokens | 137,129,435 | 282,239 |
| Unique Question Vocabulary | 1,397,929 | 19,214 |
| Unique Total Vocabulary | 1,552,766 | 20,448 |
Workflow for constructing the Italian Medical Dataset (IMD), divided into two main components: IMD-QA and IMD-MCQA. IMD-QA collects data from Italian medical forums, processing them through cleaning, tokenization, anonymization, and categorization to create a structured dataset of questions and answers. IMD-MCQA, on the other hand, integrates data from medical specialization exam simulations, applying a collection and cleaning process to generate a dataset of multiple-choice questions.
- Data Cleaning: Removal of incomplete or truncated questions, metadata (doctor signatures, timestamps), and textual inconsistencies while preserving the original medical intent.
- Text Normalization & Answer Reformulation: Answers were reformulated using Llama3-Med42-8B, a Large Language Model (LLM) fine-tuned for medical applications. The focus was on:
- Eliminating redundancy and colloquial language.
- Ensuring stylistic consistency across responses.
- Enhancing readability and grammatical accuracy.
- Anonymization: The model identified and removed personally identifiable information (PII) such as patient names, doctor names, healthcare facilities, etc.
- Data Organization: The dataset's multiple-choice questions were already structured, so the preprocessing mainly focused on standardizing the data format and ensuring consistency across entries.
The IMD-QA dataset was organized into major categories using unsupervised topic modeling. Techniques like BERTopic, UMAP, and HDBSCAN were used to group semantically similar questions into macro-categories. This approach enables flexible and interpretable categorization without rigid constraints.
| Category | # Questions |
|---|---|
| Urology, Andrology, and Male Health | 110,052 |
| Gastroenterology and Digestive Health | 104,449 |
| Mental Health | 103,893 |
| General Medicine and General Surgery | 87,789 |
| Ophthalmology, Otorhinolaryngology, Dentistry, and Pneumology | 83,710 |
| Cardiology, Circulatory System, and Hematology | 81,232 |
| Gynecology and Female Health | 65,792 |
| Orthopedics and Musculoskeletal System | 50,283 |
| Dermatology, Allergies, and Aesthetics | 49,288 |
| Neurology | 46,704 |
We welcome contributions to improve the dataset! To contribute, simply open a pull request or report issues on our issue tracker. We look forward to your improvements!
- Data Cleaning & Anonymization: Preprocessing steps have been applied to ensure privacy and data integrity while keeping the content relevant for medical applications.
- Dataset Usage: The dataset is intended for academic and research purposes only. It is not recommended for clinical decision-making or commercial use.
👨💻 This project was developed by Antonio Romano, Giuseppe Riccio, Marco Postiglione, and Vincenzo Moscato University of Naples, Federico II
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.
