README.md

---
{}
---
Below, we provide access to the datasets used in and created for the EMNLP 2022 paper [Large Language Models are Few-Shot Clinical Information Extractors](https://arxiv.org/abs/2205.12689). 

# Task #1: Clinical Sense Disambiguation
For Task #1, we use the original annotations from the [Clinical Acronym Sense Inventory (CASI) dataset](https://conservancy.umn.edu/handle/11299/137703), described in [their paper](https://academic.oup.com/jamia/article/21/2/299/723657). 
As is common, due to noisiness in the label set, we do not evaluate on the entire dataset, but only on a cleaner subset. For consistency, we use the subset defined by the filtering used in ["Zero-Shot Clinical Acronym Expansion
via Latent Meaning Cells"](https://arxiv.org/pdf/2010.02010.pdf). This results in a subset of 18,164 examples and 41 acronyms for evaluation.

We additionally use the MIMIC Reverse Substitution dataset, as created in that same paper, with further instructions available in [their repository](https://github.com/griff4692/LMC).  

# Task #2: Biomedical Evidence Extraction
For Task #2, we use the out-of-the-box high-level labels from the [PICO dataset](https://arxiv.org/abs/1806.04185) available publicly in the repository [here](https://github.com/bepnye/EBM-NLP).

# Task #3: Coreference Resolution
For Task #3, we annotated 105 snippets from the [CASI dataset](https://conservancy.umn.edu/handle/11299/137703), 5 for development and 100 for test. Each example is labeled with a singular pronoun and that pronoun's corresponding noun phrase antecedent (or antecedents).
The antecedent was annotated as the entire noun phrase (barring any dependent clauses); in cases where multiple equally valid antecedents were available, all were labeled (empirically, up to 2). 
For the purposes of evaluation, we chose the antecedent with the highest overlap to each model’s output.
To ensure nontrivial examples, the annotators excluded all examples of personal pronouns (e.g. “he”, “she”) if another person (and possible antecedent) had not yet been mentioned in the snippet.
Examples were skipped in annotation if the pronoun did not have an antecedent within the provided text snippet. 

# Task #4: Medication Status Extraction
For Task #3, we annotated 105 snippets from the [CASI dataset](https://conservancy.umn.edu/handle/11299/137703), 5 for development and 100 for test. We wanted to create a dataset of challenging examples containing a changeover in treatment. From a sample, only ∼5% of CASI snippets contained such examples. To increase the density of these examples, speeding up annotation, clinical notes were filtered with the following search terms: discont, adverse, side effect, switch, and dosage, leading to 1445 snippets. We excluded snippets that were purely medication lists, requiring at least some narrative part to be present. 
For each example, the annotators first extracted all medications. Guidelines excluded medication categories (e.g. “ACE-inhibitor”) if they referred to more specific drug names mentioned elsewhere (even if partially cut off in the snippet). For instance, only the antibiotic Levaquin was labeled in: “It is
probably reasonable to treat with antibiotics [...]. I would agree with Levaquin alone [...]”. Guidelines also excluded electrolytes and intravenous fluids as well as route and dosage information. In a second step, medications were assigned to one of three categories: active, discontinued, and neither.
Discontinued medications also contain medications that are temporarily on hold. The category neither was assigned to all remaining medications (e.g. allergies, potential medications).
The medication lists for each example were serialized as a json. 


# Task #5: Medication Attribute Extraction
For Task #5, we again annotated 105 snippets from the [CASI dataset](https://conservancy.umn.edu/handle/11299/137703), 5 for development and 100 for test.
Annotation guideline were adopted from the 2009 i2b2 medication extraction challenge (Uzuner et al., 2010) with slight modifications.
We allowed medication attributes to have multiple spans and grouped together different mentions of the the same drug (e.g. “Tylenol” and “Tylenol PM”) for the purpose of relation extraction. 
The annotation list for each example was serialized as a json. 

# Citations
When using our annotations for tasks #3-5, please cite our paper, as well as the papers from which the underlying text originated. 

```
@inproceedings{agrawal2022large,
    title={Large Language Models are Few-Shot Clinical Information Extractors},
    author={Monica Agrawal and Stefan Hegselmann and Hunter Lang and Yoon Kim and David Sontag},
    booktitle = {Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing},
    year={2022},
    url_Paper = {https://arxiv.org/pdf/2205.12689.pdf}
}
```

```
@article{moon2014sense,
  title={A sense inventory for clinical abbreviations and acronyms created using clinical notes and medical dictionary resources},
  author={Moon, Sungrim and Pakhomov, Serguei and Liu, Nathan and Ryan, James O and Melton, Genevieve B},
  journal={Journal of the American Medical Informatics Association},
  volume={21},
  number={2},
  pages={299--307},
  year={2014},
  publisher={BMJ Publishing Group BMA House, Tavistock Square, London, WC1H 9JR}
}
```

# Licensing
The annotations added by our team fall under the MIT license, but the CASI dataset itself is subject to its own licensing. 



---
license: other
---