README.md

metadata

license: cc-by-4.0
pretty_name: >-
  Topic model data from Finding Long-COVID: temporal topic modeling of
  electronic health records from the N3C and RECOVER programs

Note: very small numbers are rounded to 0 in the HuggingFace dataset viewer.

These data represent medical concept probabilities for 300 topics generated via Latent Direchlet Allocation applied to 387M electronic health record conditions for 7.9M patients as described in Finding Long-COVID: temporal topic modeling of electronic health records from the N3C and RECOVER programs . Data were quality filtered and cleaned prior to modeling, including removal of COVID-19 and MIS-C as confounders (see publication). Topic T-23 was most strongly associated with Long-COVID across all demographics.

If you use these data please cite the publication above.

Terms are encoded as OMOP CDM standard concept_id values; see details at OHDSI and OHDSI ATHENA.

Columns:

• topic_name: topics are named T-1 to T-300, in order of weighted usage; also included in topic name are U (aggregate usage of topic in data), H (a measure of topic usage uniformity across data contributing sites), and C (a z-score normalized coherence value indicating relative topic quality)
• concept_name: human-readable name of the OMOP CDM concept
• concept_id: OMOP CDM standard concept ID
• term_weight: the probability of the concept_id being generated by the topic topic_name
• relevance: A measure of topic-relative weight - positive values indicate concepts more highly weighted in the topic than over all data, negative values less.