Dataset Card: SIMORD (Simulated Medical Order Extraction Dataset)

Description

Medical order extraction involves identifying and structuring various medical orders —such as medications, imaging studies, lab tests, and follow-ups— based on doctor-patient conversations. Previous efforts have focused on extracting entities and relations from clinical texts.

This dataset seeks to encourage the developement of effective solutions for improving clinical documentation, reducing the burden on providers, and ensuring critical patient information is accurately captured from long conversations.

The input dialogues are sourced from a combination of existing conversational datasets (e.g., ACI-Bench [1], PriMock57 [2]), and structured lists of medical orders are created by medical annotators.

Dataset Summary

• Name: SIMORD
• Full name / acronym: SIMulated ORDer Extraction
• Purpose / use case:
  SIMORD is intended to support research in extracting structured medical orders (e.g. medication orders, lab orders) from doctor-patient consultation transcripts.
• Version: As released with the EMNLP industry track paper (2025)
• License / usage terms: CDLA-2.0-permissive
• Contact / Maintainer: jcorbeil@microsoft.com

Building the dataset

Method 1: HF datasets

1. Make sure you have datasets==3.6.0 or less, otherwise builder is not supported in recent versions.
2. Git clone and install requirements from https://github.com/jpcorb20/mediqa-oe
3. Add mediqa-oe to python path PYTHONPATH=$PYTHONPATH:/mypath/to/mediqa_oe (UNIX).
4. Run load_dataset("microsoft/SIMORD", trust_remote_code=True), which will merge transcripts from ACI-Bench and Primock57 repos into the annotation files.

Method 2: GitHub script

Follow the steps in https://github.com/jpcorb20/mediqa-oe to merge transcripts from ACI-Bench and Primock57 into the annotation files provided in the repo.

Data Fields / Format

Input fields:

• transcript (dict of list): the doctor-patient consultation transcript as dict of three lists using those keys:
  • turn_id (int): index of that turn.
  • speaker (str): speaker of that turn DOCTOR or PATIENT.
  • transcript (str): line of that turn.

Output fields:

• A JSON (or list) of expected orders
• Each order object includes at least:
  • order_type (e.g. “medication”, “lab”)
  • description (string) — the order text (e.g. “lasix 40 milligrams a day”)
  • reason (string) — the clinical reason or indication for the order
  • provenance (e.g. list of token indices or spans) — mapping back to parts of the transcript

Splits

• train: examples for in-context learning or fine-tuning.
• test1: test set used for the EMNLP 2025 industry track paper. Also, previously named dev set for MEDIQA-OE shared task of ClinicalNLP 2025.
• test2: test set for MEDIQA-OE shared task of ClinicalNLP 2025.

Citation

If you use this dataset, please cite:

@article{corbeil2025empowering,
  title={Empowering Healthcare Practitioners with Language Models: Structuring Speech Transcripts in Two Real-World Clinical Applications},
  author={Corbeil, Jean-Philippe and Abacha, Asma Ben and Michalopoulos, George and Swazinna, Phillip and Del-Agua, Miguel and Tremblay, Jerome and Daniel, Akila Jeeson and Bader, Cari and Cho, Yu-Cheng and Krishnan, Pooja and others},
  journal={arXiv preprint arXiv:2507.05517},
  year={2025}
}

References

[1] Aci-bench: a Novel Ambient Clinical Intelligence Dataset for Benchmarking Automatic Visit Note Generation. Wen-wai Yim, Yujuan Fu, Asma Ben Abacha, Neal Snider, Thomas Lin, Meliha Yetisgen. Nature Scientific Data, 10, 586 (2023).

[2] PriMock57: A Dataset Of Primary Care Mock Consultations. Alex Papadopoulos Korfiatis, Francesco Moramarco, Radmila Sarac, and Aleksandar Savkov. 2022. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 588–598, Dublin, Ireland.