kbc-pomr

Code for the paper "Knowledge Base Completion for Constructing Problem-Oriented Medical Records" at MLHC 2020

Data

All annotations can be found at data/all.csv. Each row lists a problem, a relation type (also the data type of the target), and the target code, along with the annotated label (1 = negative, 2 = positive).

In the data/ directory, we also have the train, dev, and test splits for both experiments conducted in the paper. *_probs.csv files contains data splits, separated by problem type (Table 3), and *_rand.csv files contains data splits, separated at random (Table 2).

We also provide:

• data/med_may_treat.csv - An auxiliary lookup to find the SNOMED/ICD diagnosis codes that an RxNorm code may be related to, which we constructed by going through NDF-RT's "MayTreat" and "MayPrevent" relations
• data/problem_codes_all.csv - A file with our problem definitions
• data/site_icd9_relative_freqs.csv - A file with the relative frequencies of ICD-9 codes computed from our EHR dataset, to properly initialize problem embeddings
• intersect_*.txt: the lists of codes for each data type that we evaluate on, which we constructed by taking the intersection with the set of site-specific codes.
• vocab.txt - the vocabulary used (site-specific codes censored with X's)
• embeddings/claims_codes_hs_300.txt and embeddings/claims_cuis_hs_300.txt - the code and CUI embeddings from Choi et al

Reproduction of results

Download open-source embeddings

First, download and extract (with gunzip) the embeddings for codes (here) and CUIs (here) from prior work, and put the files in the embeddings/ directory.

Environment setup

To set up the proper dependencies using conda, run:

conda create -n POMR python=3.7

conda activate POMR

pip install -r requirements.txt

Reproducing experiments

The jupyter notebook Reproduction.ipynb gives full instructions to reproduce the results from the paper, specifically line 4 ("Choi et al") in Table 2 and lines 1 ("Ontology baseline") and 5 ("Choi et al") in Table 3.

At a high level, the steps are:

• Construct RxNorm-to-CUI lookup using UMLS, so we can use Choi et al's medication embeddings
• Pre-compute problem and target embeddings to use to initialize models.
• Train on the held-out triplets data splits (*_rand.csv) to reproduce Table 2
• Train on the held-out problems data splits (*_newprobs.csv) to reproduce Table 3

Citation

If you use this repository, please cite our paper:

@inproceedings{mullenbach2020knowledge,
  title={Knowledge Base Completion for Constructing Problem-Oriented Medical Records},
  author={Mullenbach, James and Swartz, Jordan and McKelvey, T Greg and Dai, Hui and Sontag, David},
  booktitle={Machine Learning for Healthcare Conference},
  year={2020}
}