Clinical Concept Co-Occurrence Matrices from 20 Million Notes

Offering co-occurrence matrices quantifying pairwise mentions of 3 million terms mapped to 1 million clinical concepts. It was computed from the raw text of 20 million clinical notes spanning 19 years of data from Stanford Hospitals and Clinics.

Use Cases

• Compute Bayesian conditional probabilities between medical concepts using the co-occurrence matrix to assess drug-drug interaction patterns.
• Quantify the relatedness of diseases and procedures for epidemiological studies of comorbidity based on clinical concept mentions.
• Generate association rules from the co-frequencies of drugs and devices to inform clinical decision support systems.
• Calculate statistical measures like relative risks and odds ratios for drug-disease relationships using the pairwise mention data.

Strengths

• Derived from a large corpus of 20 million clinical notes.
• Covers a long temporal span of 19 years of clinical data.
• Maps a vast vocabulary of 3 million terms onto 1 million standardized clinical concepts.
• Quantifies pairwise co-occurrences, enabling direct computation of various statistical measures.

Limitations

• Data originates from a single institution (Stanford), limiting geographic and institutional generalizability.
• The underlying clinical notes are from 2000-2019, so the data may not reflect recent medical practices or terminology.
• The co-occurrence matrices are a derived, aggregated view; the raw clinical note text is not provided.

Provenance

Source

Clinical notes from Stanford Hospitals and Clinics.

Collection Method

Co-frequencies computed via a parallelized annotation, hashing, and counting pipeline applied to clinical narrative text.

Time Range

19 years of data (span not explicitly defined, but notes cover up to 2019).

Freshness

Last updated in June 2020.

Geography

Stanford Hospitals and Clinics, United States.