PMC-Patients Meta Data

## PMC-Patients Meta Data

Meta data for PMC-Patients that might facilitate reproduction or usage of our dataset, consisting of the following files (most of which can be derived from our main files above).

### PMIDs.json

PMIDs of articles from which PMC-Patients are extracted.
List of string, length 140,897.

### train_PMIDs.json & dev_PMIDs.json & test_PMIDs.json & human_PMIDs.json

PMIDs of articles in training / dev / test split.
List of string.

### train_patient_uids.json & dev_patient_uids.json & test_patient_uids.json & human_patient_uids.json

Patient_uids of notes in training / dev / test split.
List of string.

### patient2article_relevance.json

Full patient-to-article dataset.
A dict where the keys are `patient_uid` of queries and each entry is a list of `PMID`, representing articles relevant to the query.

The 3-point relevance can be obtained by checking whether the `PMID` is in `PMIDs.json`.

### patient2patient_similarity.json

Full patient-to-patient similarity dataset.
A dict where the keys are `patient_uid` of queries and each entry is a list of `patient_uid`, representing similar patients to the query.

The 3-point similarity can be obtained by checking whether the similar patient share the `PMID` (the string before '-' in `patient_uid`) with the query patient.

### PMID2Mesh.json

Dict of PMIDs to MeSH terms of the article.

### MeSH_Humans_patient_uids.json

`patient_uid` of the patients in PMC-Patients-Humans (extracted from articles with "Humans" MeSH term).
List of string.

### PMC-Patients_citations.json

Citations for all articles we used to collect our dataset.
A dict where the keys are `patient_uid` and each entry is the citation of the source article.

### human_PMIDs.json

PMIDs of the 500 randomly sampled articles for human evaluation.
List of string.

### PMC-Patients_human_eval.json

Expert annotation results of the 500 articles in `human_PMIDs.json`, including manually annotated patient note, demographics, and relations of the top 5 retrieved articles / patients.
List of dict, and the keys are almost identical to `PMC-Patients.json`, with the exception of `human_patient_id` and `human_patient_uid`.

The relational annotations are different from automatic ones. They are strings indicating on which dimension(s) are the patient-article / patient-patient pair relevant / similar.
"0", "1", "2", and "3" represent "Irrelevant", "Diagnosis", "Test", "Treatment" in ReCDS-PAR, and represent "Dissimilar", "Features", "Outcomes", "Exposure" in ReCDS-PPR.
Note that a pair can be relevant / similar on multiple dimensions at the same time.

### PAR_PMIDs.json

PMIDs of the 11.7M articles used as PAR corpus.
List of string.