Elixhauser comorbidity profile (31 domains, van Walraven score)

Flags the 31 Elixhauser comorbidity domains (Quan et al. 2005 coding algorithms) in the look-back window before each cohort row's index date and scores them with the van Walraven 2009 weights. It is the R counterpart of Python omop_etl.extract_elixhauser_comorbidities(): both read the same curated code table, system.file("extdata", "comorbidity", "elixhauser_codes.csv", package = "omopduckdb") (sources, weights, hierarchy rules and known deviations are in PROVENANCE.md next to it), build the same SQL, and return the same columns and values.

Usage

extract_elixhauser_comorbidities(
  con,
  cohort_table = "cohort",
  cohort_id = NULL,
  person_col = "subject_id",
  index_date_col = "cohort_start_date",
  lookback_days = 365,
  source_and_standard = TRUE,
  hierarchy_adjusted = TRUE,
  exclude_index_diagnoses = TRUE,
  fallback_on_unmapped = FALSE,
  return_mapping_fidelity = FALSE
)

Arguments

con

Active DBI connection to DuckDB holding condition_occurrence and the cohort table (read-only is fine). concept_ancestor is looked up as concept_ancestor (main, a TEMP view or the connection's search_path, e.g. after omop_connect()), then as central_vocab.main.concept_ancestor.

cohort_table

Table or view with one row per index event (default "cohort"); may be schema-qualified.

cohort_id

Optional integer. Keep only rows with cohort_definition_id = cohort_id. If the table has no cohort_definition_id column a warning is raised and every row is profiled.

person_col

Person id column of cohort_table, joined to condition_occurrence.person_id. It must exist: columns are never guessed.

index_date_col

Index date column of cohort_table; must exist. It is cast to DATE to place the window.

lookback_days

Window [index - lookback_days, index): condition_start_date strictly before the index date and no more than lookback_days days earlier. NULL means all history before the index date.

source_and_standard

Also match condition_concept_id against the domain's SNOMED ancestors and their descendants (default TRUE). If concept_ancestor cannot be resolved on the connection (or is empty) a warning is raised and only the source side is used. FALSE uses the source side only, without a warning.

hierarchy_adjusted

Apply the hierarchy rules to the score and the condition count (default TRUE). FALSE gives the plain weighted sum and the plain count of raw flags.

exclude_index_diagnoses

Bedside-scoring guard (train/serve skew, default TRUE): diagnoses are coded by medical coders days to weeks after discharge, so a model trained on them cannot be scored at discharge. When TRUE, (a) condition records whose visit_occurrence_id equals the cohort row's visit_occurrence_id are never counted (when both tables have that column; cohorts from define_study_cohort() do), and (b) an index_date_col that looks like a discharge / end date is an error, since a window ending at discharge contains the index stay. The window is always [index - lookback_days, index).

fallback_on_unmapped

Logical; if TRUE, falls back to source-value matching if standard concept ancestor is unavailable (default FALSE).

return_mapping_fidelity

Logical; if TRUE, attaches mapping fidelity diagnostics attribute (default FALSE).

Value

A data.frame with the columns, in order: person_col, index_date_col, the 31 elix_<domain> flags (integer 0/1: chf, arrhythmia, valvular, pulm_circ, pvd, htn_uncomp, htn_comp, paralysis, neuro_other, copd, dm_uncomp, dm_comp, hypothyroid, renal_failure, liver_disease, pud, hiv, lymphoma, mets, solid_tumor, rheumatic, coagulopathy, obesity, weight_loss, fluid_electrolyte, blood_loss_anemia, deficiency_anemia, alcohol_abuse, drug_abuse, psychoses, depression), elix_van_walraven_score (integer; the weights run from -7 to 12, so it can be negative) and elix_total_conditions (integer count of flagged domains).

Details

The result has one row per DISTINCT (person, index date) of the cohort table, ordered by person then index date. Persons without a qualifying condition still appear, with every flag 0 and a score of 0.

A condition record counts for a domain when its condition_source_value, upper-cased with . and whitespace removed, STARTS WITH one of the domain's ICD-10-CM / ICD-9-CM prefixes (I50 matches I50.9 and i 50.9; 4280 matches 428.0), or (source_and_standard) its condition_concept_id is a SNOMED ancestor of the domain or a descendant of one in concept_ancestor. Only records dated strictly before the index date are used ([index - lookback_days, index)), so nothing at or after the index leaks in.

The profile is one statement of CTEs: nothing is created, so it works on read-only connections and leaves no objects behind.

Hierarchy. With hierarchy_adjusted = TRUE, dm_comp supersedes dm_uncomp, mets supersedes solid_tumor and htn_comp supersedes htn_uncomp in the score and in the condition count; the 0/1 domain columns always stay the raw flags. Only the mets rule changes the score (12 instead of 16 together with a solid tumour); the diabetes and hypertension rules change the count. The standard-concept side of dm_uncomp is the set of diabetes types, which includes every complicated form, so its raw flag is only meaningful after the hierarchy is applied.

Prefixes are matched against any source value, whatever its coding system, so an ICD-9-CM V prefix can match an ICD-10-CM transport-accident code and the ICD-10-CM prefixes E00-E03, E86, E87 and E890 can match ICD-9-CM external-cause E codes (PROVENANCE.md lists them); profile data whose coding system you know.

See also

extract_charlson_index()

Examples

con <- DBI::dbConnect(duckdb::duckdb())
#> duckdb keeps downloaded extensions and secrets in a temporary directory:
#> ℹ /tmp/RtmpFV3zps/duckdb
#> This is removed when the R session ends.
#> • Extensions are re-downloaded each session.
#> • Secrets are lost.
#> ℹ Run duckdb(shared_home = TRUE) (or create ~/.duckdb) to keep them (suitable for most users).
#> ℹ Run duckdb(shared_home = FALSE) to accept the temporary directory (and silence this message).
#> ℹ See ?duckdb_storage for details and alternatives.
DBI::dbExecute(con, "CREATE TABLE condition_occurrence AS
  SELECT 1 AS person_id, 0 AS condition_concept_id, DATE '2021-05-01' AS condition_start_date,
         'I50.9' AS condition_source_value")
#> [1] 1
DBI::dbExecute(con, "CREATE TABLE cohort AS
  SELECT 1 AS subject_id, DATE '2021-06-01' AS cohort_start_date")
#> [1] 1
extract_elixhauser_comorbidities(con, source_and_standard = FALSE)[
  , c("subject_id", "elix_chf", "elix_van_walraven_score")]
#>   subject_id elix_chf elix_van_walraven_score
#> 1          1        1                       7
DBI::dbDisconnect(con, shutdown = TRUE)