This repository hosts the work detailed in our paper, BIT.UA at MedProcNer: Discovering Medical Procedures in Spanish Using Transformer Models with MCRF and Augmentation.
- Tiago Almeida - [email protected] - ORCID: 0000-0002-4258-3350
- Richard A. A. Jonker - [email protected] - ORCID: 0000-0002-3806-6940
- Roshan Poudel - [email protected] - ORCID: 0000-0001-6450-023X
- Jorge M. Silva - [email protected] - ORCID: 0000-0002-6331-6091
- Sérgio Matos - [email protected] - ORCID: 0000-0003-1941-3983
We use the MedProcNER corpus in our study. The MedProcNER corpus is a collection of 1,000 clinical case reports in Spanish annotated with clinical procedure mentions and normalized to SNOMED CT. You can access the dataset here.
- Gold Standard data (includes train and test set, as well as gazetteers)
- Annotation Guidelines (Spanish)
- HF_CACHE : cache directory for downloading models from Hugging Face
- dataset : folder for MedProcNER dataset
- medprocner-submission : folder containing submission for all subtasks (results)
- subtask-1 : code for subtask 1 NER
- subtask-2 : code for subtask 2, entity linking to concepts
- subtask-3 : code for subtask 3, indexing the concepts
- trained_models : folder containing trained models
Prerequisites: Python version 3.9 (minimum); We used Python version 3.10
git clone https://github.com/ieeta-pt/MedProcNER.git
chmod +x setup.sh
./setup.sh # installs environments and requirement dependenciesThe make_runs.sh scripts are available in the respective subtask directories. They're shell scripts which execute a series of commands for replicating the results we obtained in our study. Here's how to use them and what they do:
To replicate the results of Subtask-1, navigate to the subtask-1 directory and run make_runs.sh:
cd subtask-1
./make_runs.shThis script performs the following steps:
- Downloads the models used for this subtask.
- Runs the downloaded models on the provided dataset.
- Combines the results of different model runs to form an ensemble.
- Converts the results to the MedProcNER format for submission.
To replicate the results of Subtask-2, navigate to the subtask-2 directory and run make_runs.sh:
cd ../subtask-2
./make_runs.shThis script performs the following steps:
- Downloads the models used for this subtask.
- Creates a semantic index for the SNOMED CT medical terminology system.
- Performs a semantic search on the indexed terms using the results from the model runs.
- Converts the results to the MedProcNER format for submission.
To replicate the results of Subtask-3, navigate to the subtask-3 directory and run make_runs.sh:
cd ../subtask-3
./make_runs.shThis script performs the following steps:
- Get the concepts from each document and concider all of them as the document index.
- The resulting index list represents the results for Subtask-3 and is ready for submission.
By running these scripts, you can replicate our submissions for each subtask.
Check the README in the subtask-1 folder to learn how to train the NER models.
Comparison of System Performances on Subtasks with Best Competition Results:
| System | Models Used | Aug. | Data | Subtask-1 F1 | Subtask-2 F1 | Subtask-3 F1 |
|---|---|---|---|---|---|---|
| system-0 | 5 x lcampillos-dense | UNK | no val | 79.46 | 31.53 | 35.67 |
| system-1 | 5 x lcampillos-dense | UNK | full | 79.24 | 31.32 | 35.42 |
| system-2 | 30 x lcampillos-bilstm | mixed | mixed | 78.81 | 31.10 | 35.37 |
| system-3 | 20 x plantl-mixed | mixed | no val | 79.23 | 31.66 | 35.98 |
| system-4 | All plantl and lcampilos (91) | mixed | mixed | 79.85 | 31.68 | 35.85 |
| Best | - | - | - | 79.85 | 57.07 | 62.42 |
For any issue you encounter with the repository or code, please report it to the Issues section.
If you find this helpful work and use it in your own research, please cite our paper as follows:
@inproceedings{bituaMedprocner,
title={ {BIT.UA} at {M}ed{P}roc{N}er: Discovering Medical Procedures in Spanish Using Transformer Models with {MCRF} and Augmentation},
author={Almeida, Tiago and Jonker, Richard A. A. and Poudel, Roshan and Silva, Jorge M. and Matos, Sérgio},
booktitle={Working Notes of CLEF 2023 - Conference and Labs of the Evaluation Forum},
year={2023}
}