omop-duck-db

Tools for building an OMOP Common Data Model (CDM) v5.4 database in DuckDB, loading the OHDSI Athena vocabulary, and mapping PCORnet-format source data into the CDM. Provided as both an R package and a Python package, with parallel ETL implementations that share the same schema and concept-mapping SQL, so a human (or an AI agent) picking either language gets the same result.

Scope

This repo provides general-purpose mapping/ETL tooling and a small fabricated sample fixture for testing – it does not itself ingest, store, or process any protected or identifiable data. Real-world use against production source data happens outside this repo, against your own data infrastructure.

Layout

inst/extdata/5.4/duckdb/*.sql — OMOP CDM v5.4 DDL, generated once via the CommonDataModel R package. This is the schema source of truth for both the R and Python ETL paths — neither re-declares CREATE TABLE by hand.
inst/sql/mapping_macros.sql — DuckDB SQL macros for source-code → standard-concept resolution, shared by both ETLs.
R/ — the R package: schema build, vocabulary load/refresh, PCORnet ETL, Google Drive snapshot sync, BigQuery live connector.
inst/scripts/ — thin command-line wrappers around the R package’s functions, for scripting/automation use without writing R code.
python/omop_etl/ — the Python package (build_omop_cdm.py), mirrors the R package table-for-table.
tests/fixtures/pcornet_sample/ — small fabricated sample rows used to smoke-test both ETLs against each other.
config/ — non-secret config templates (*.yml.example). Copy to *.yml and fill in locally; the real files are gitignored.

The generated database file and any vocabulary download archive are not committed to git — they’re large binaries. See “Large files” below.

Installation

# install.packages("remotes")
remotes::install_github("jkylearmstrong/omop-duck-db")

Python:

pip install omop-duck-db
# or for ML dependencies: pip install 'omop-duck-db[ml]'

Developer Note: `renv` Workflow

This repository includes an .Rprofile that activates renv automatically (renv/activate.R). - If working within the project’s isolated environment: Run renv::restore() to populate local package libraries. - If using pre-installed system R packages: Set RENV_CONFIG_AUTOLOADER = FALSE in .Renviron or invoke Rscript --no-init-file so standalone scripts do not fail looking for unpopulated local renv libraries.

Example

A schema builds into a fresh, empty DuckDB file with a single call:

library(omopduckdb)

db_path <- tempfile(fileext = ".duckdb")
build_schema(db_path)
#> Schema built at C:\Users\jkyle\AppData\Local\Temp\RtmpoJ2PVA\file6a0864f2dde.duckdb

con <- DBI::dbConnect(duckdb::duckdb(), db_path, read_only = TRUE)
DBI::dbListTables(con)[1:10]
#>  [1] "care_site"            "cdm_source"           "cohort"              
#>  [4] "cohort_definition"    "concept"              "concept_ancestor"    
#>  [7] "concept_class"        "concept_relationship" "concept_synonym"     
#> [10] "condition_era"
DBI::dbDisconnect(con, shutdown = TRUE)

From there, load_vocabulary() loads a vocabulary download, and etl_pcornet() maps a PCORnet-format source directory into the CDM tables just created — see “Building the database” below for the full workflow, or inst/scripts/ for command-line equivalents of each step.

Building the database

Generate the CDM v5.4 schema and load the vocabulary:

library(omopduckdb)
build_schema("omop_cdm.duckdb")
load_vocabulary("path/to/vocabulary_download", "omop_cdm.duckdb")

Map a source directory into the CDM, with either implementation (they read the same source layout and produce equivalent output):

etl_pcornet("path/to/source_extract", "omop_cdm.duckdb")

python -m omop_etl.build_omop_cdm --source-dir path/to/source_extract

Command-line equivalents of the R functions live in inst/scripts/, e.g. Rscript inst/scripts/etl_pcornet.R --source-dir path/to/source_extract.

Refreshing the vocabulary

From a new download: load_vocabulary("path/to/new_download", "omop_cdm.duckdb"). Each table is truncated and reloaded inside its own transaction with a row-count sanity check, so a bad/partial file can only fail that one table (rolled back, left untouched), not half-clobber the rest of the vocabulary.
Live, from BigQuery (e.g. a public or your own OMOP-shaped dataset): load_from_bigquery(billing_project = "...", dataset = "...", tables = c("concept", "concept_ancestor", "concept_relationship")). A GCP project ID is required for query billing/quota even against public datasets (set one in config/bigquery.yml, copied from config/bigquery.yml.example).

First-time auth: the BigQuery OAuth flow needs an interactive R session to open a browser — it fails silently if run non-interactively before a token exists. Run this once, in an interactive session (not Rscript), and approve the browser prompt:
```
bigrquery::bq_auth()
```
The resulting token is cached and reused by later non-interactive runs.

Transparent Central Vocabulary & Single-Store Architecture (`omop_connect`)

Instead of duplicating 10–15 GB of Athena vocabulary tables inside every site-specific CDM database, omop_connect() attaches a shared central vocabulary database read-only (central_vocab) and automatically injects DuckDB’s search_path = 'main,central_vocab'. All queries seamlessly resolve concept, concept_ancestor, and vocabulary macros without schema qualification:

library(omopduckdb)

# Auto-discovers sibling central_vocabulary.duckdb or connects with explicit path
con <- omop_connect("omop_cdm.duckdb")
DBI::dbGetQuery(con, "SELECT COUNT(*) FROM concept;")

In Python:

from omop_etl import omop_connect

con = omop_connect("omop_cdm.duckdb")
con.execute("SELECT COUNT(*) FROM concept").fetchone()

Custom mapping and retroactive remapping

When standard Athena vocabularies lack mappings for site-specific or unmapped codes:

Audit unmapped codes:

export_unmapped_codes("omop_cdm.duckdb", "condition_occurrence", "unmapped_conditions.csv")

Import custom crosswalks or Usagi exports:

# Load custom source_to_concept_map CSV
import_source_to_concept_map("mappings.csv", "omop_cdm.duckdb")

# Or import OHDSI Usagi review export CSV
import_usagi_mappings("usagi_reviewed.csv", "omop_cdm.duckdb")

Retroactively remap populated CDM tables:

# Preview changes without modifying data (dry run)
remap_cdm_table("condition_occurrence", "omop_cdm.duckdb", dry_run = TRUE)

# Apply remapping across all tables and rebuild condition/drug eras
remap_all("omop_cdm.duckdb")

Equivalent functions are available in Python via from omop_etl import remap_all, import_usagi_mappings.

Study Cohorts, Phenotyping & Clinical Pipeline

The package provides an end-to-end epidemiological pipeline for observational studies, featuring flexible index stay definition, validated comorbidity profiling, physiologic range sanitization, Athena concept hierarchy rollups, LOINC laboratory consolidation, and publication Table 1 generation.

1. Unified Study Cohort Engine (`define_study_cohort`)

Generalizes acute inpatient, emergency department, and outpatient index cohorts with customizable wash-in windows, right-censoring verification, and outcome definitions:

library(omopduckdb)

con <- DBI::dbConnect(duckdb::duckdb(), "omop_cdm.duckdb")

# Build general study cohort (or specialized build_readmission_cohort / build_end_of_life_cohort)
cohort_df <- define_study_cohort(
  con,
  visit_types = "inpatient",
  washin_days = 365,
  followup_days = 30,
  min_age = 18,
  min_los_days = 1,
  sampling_rule = "first",
  require_verified_followup = TRUE,
  target_table = "cohort"
)

In Python:

from omop_etl import define_study_cohort

cohort_df = define_study_cohort(
    con,
    visit_types="inpatient",
    washin_days=365,
    followup_days=30,
    min_age=18,
    min_los_days=1,
    sampling_rule="first",
    require_verified_followup=True,
    target_table="cohort"
)

2. Comorbidity Profiling: 31 Elixhauser Domains & Charlson Index

Computes validated comorbidity profiles in one set-based DuckDB query over the cohort with zero object leakage on read-only connections. Automatically evaluates both ICD-9/10 prefixes (condition_source_value) and SNOMED descendants (condition_concept_id):

# Extract 31 Elixhauser domains + composite van Walraven score
elix_df <- extract_elixhauser_comorbidities(con, cohort_table = "cohort", lookback_days = 365)

# Extract 17 Quan-Charlson categories + Charlson Comorbidity Index
cci_df <- extract_charlson_index(con, cohort_table = "cohort", lookback_days = 365)

In Python:

from omop_etl import extract_elixhauser_comorbidities, extract_charlson_index

elix_df = extract_elixhauser_comorbidities(con, cohort_table="cohort", lookback_days=365)
cci_df = extract_charlson_index(con, cohort_table="cohort", lookback_days=365)

3. Physiologic Measurement Sanitization & Outlier Winsorization (`sanitize_measurements`)

Cleans laboratory and vital measurements against 136 standard OHDSI DataQualityDashboard biological plausibility limits or statistical IQR/z-score winsorization:

# Clamps out-of-bounds measurements to DQD biological plausibility limits
clean_labs <- sanitize_measurements(
  con,
  method = "dqd_biologic_limits",
  action = "clamp"
)

In Python:

from omop_etl import sanitize_measurements

clean_labs = sanitize_measurements(
    con,
    method="dqd_biologic_limits",
    action="clamp"
)

4. Build Inpatient Readmission Cohort (`build_readmission_cohort`)

Constructs index inpatient stays (visit_concept_id = 9201) among adult patients ( $\ge 18$ ) with length of stay $\ge 1$ day, discharged alive, and verified 30-day follow-up (requiring either a readmission or confirmed clinical contact $\ge 30$ days post-discharge to prevent right-censoring):

library(omopduckdb)

con <- DBI::dbConnect(duckdb::duckdb(), "omop_cdm.duckdb")

# Build cohort into OMOP cohort table (cohort_definition_id = 1)
cohort_df <- build_readmission_cohort(
  con,
  washin_days = 365,
  followup_days = 30,
  index_rule = "random",
  random_seed = 42,
  require_verified_followup = TRUE,
  target_table = "cohort"
)

In Python:

import duckdb
from omop_etl import build_readmission_cohort

con = duckdb.connect("omop_cdm.duckdb")
cohort_df = build_readmission_cohort(
    con,
    washin_days=365,
    followup_days=30,
    index_rule="random",
    random_seed=42,
    require_verified_followup=True,
    target_table="cohort"
)

5. Extract Temporal Utilization & Demographic Features

Aggregates windowed baseline healthcare utilization (IP, ED, OP encounter counts, days since prior visit, prior 30-day readmissions in lookback), index length of stay, discharge disposition, and demographics:

temporal_features <- extract_temporal_features(
  con,
  cohort_table = "cohort",
  cohort_definition_id = 1,
  lookback_days = 365
)

In Python:

from omop_etl import extract_temporal_features

temporal_features = extract_temporal_features(
    con,
    cohort_table="cohort",
    cohort_definition_id=1,
    lookback_days=365
)

6. Athena Concept Hierarchy Rollup & Consolidated Measurements

Traverse Athena’s concept_ancestor table to aggregate medication or condition classes, and extract consolidated lab panels and physical vitals using acute pre-discharge aggregation strategies:

# Aggregate medication exposure using Athena ancestor traversal
med_features <- aggregate_concept_sets(
  con,
  cohort_table = "cohort",
  cohort_definition_id = 1,
  domain = "drug",
  concept_sets = DEFAULT_MEDICATION_CONCEPTS,
  lookback_days = 365
)

# Extract acute pre-discharge laboratory and vital measurements
lab_features <- extract_measurements(
  con,
  cohort_table = "cohort",
  cohort_definition_id = 1,
  loinc_map = DEFAULT_LAB_LOINCS,
  strategy = "last_before_discharge"
)

In Python:

from omop_etl import (
    aggregate_concept_sets,
    extract_measurements,
    DEFAULT_MEDICATION_CONCEPTS,
    DEFAULT_LAB_LOINCS
)

med_features = aggregate_concept_sets(
    con,
    cohort_table="cohort",
    cohort_definition_id=1,
    domain="drug",
    concept_sets=DEFAULT_MEDICATION_CONCEPTS,
    lookback_days=365
)

lab_features = extract_measurements(
    con,
    cohort_table="cohort",
    cohort_definition_id=1,
    loinc_map=DEFAULT_LAB_LOINCS,
    strategy="last_before_discharge"
)

7. Stratified Table 1 Generation & Completeness Matrix

Generates baseline descriptive tables (continuous: Mean +/- SD, Median [IQR]; categorical: N [%]) stratified by outcome (e.g., readmitted_30d), calculating Standardized Mean Differences (SMD) and statistical tests, along with a Table 1b completeness audit:

table1_out <- generate_table1(
  patient_df,
  strata = "readmitted_30d",
  continuous_vars = c("age_at_admission", "los_days", "prior_ed_visits"),
  cat_vars = c("gender", "race", "tobacco_status")
)

# Baseline summary with SMD and p-values
print(table1_out$table1)

# Missingness and completeness audit
print(table1_out$table1b_missingness)

In Python:

from omop_etl import generate_table1

t1_df, t1b_df = generate_table1(
    patient_df,
    strata="readmitted_30d",
    continuous_vars=["age_at_admission", "los_days", "prior_ed_visits"],
    cat_vars=["gender", "race", "tobacco_status"]
)

8. Benchmark Reconciliation & Validation Harness

Validates prospective OMOP Table 1 statistics against benchmark datasets with configurable absolute tolerance thresholds, flagging distribution drift and missing features:

reconciliation <- validate_table1_reconciliation(
  omop_table1 = table1_out$table1,
  source_table1 = benchmark_table1,
  tolerance = 0.05
)

if (!reconciliation$passed) {
  warning("Table 1 discrepancy detected!")
  print(reconciliation$discrepancies)
}

In Python:

from omop_etl import validate_table1_reconciliation

reconciliation = validate_table1_reconciliation(
    omop_table1=t1_df,
    source_table1=benchmark_table1,
    tolerance=0.05
)
print("Reconciliation Passed:", reconciliation["passed"])

ML bridge: omop-learn, scikit-learn, SARD & OHDSI PLP / DeepPLP

omop-duck-db connects an OMOP-on-DuckDB database to three modelling ecosystems without ad-hoc SQL or per-patient query loops. Every extractor runs as one set-based DuckDB query over the whole cohort, and every window is strictly before the index date by default ([index - lookback_days, index)), so nothing at or after the index leaks into the features. Concept id 0 (“No matching concept”) is always dropped.

You use	Entry point
scikit-learn / glmnet / LASSO	`extract_sparse_concept_matrix()` (Python: `scipy.sparse.csr_matrix`; R: `Matrix::dgCMatrix`)
SARD / sequence models	`extract_sard_visit_tensors()` (padded `(N, max_nvisits, max_visit_len)`)
omop-learn (`OMOPDataset`, its torch / sparse / windowed datasets and models)	`DuckDBBackend`, a DuckDB implementation of omop-learn’s backend interface
OHDSI PatientLevelPrediction / DeepPatientLevelPrediction	`plp_database_details()`, `hades_preflight()`, `as_plp_data()`
tidymodels / in-database scoring	`as_tidymodels_data()`, `model_concepts()`, `materialize_concept_features()`

1. Sparse concept matrix and SARD visit tensors

Both take a cohort parquet (one row per index event; person id and index date columns are auto-detected, an outcome column is carried through) and return arrays whose row i is parquet row i.

import duckdb
from omop_etl import extract_sparse_concept_matrix, extract_sard_visit_tensors

con = duckdb.connect("derived/omop_duckdb/omop_Temple.duckdb", read_only=True)
con.execute("ATTACH 'derived/omop_duckdb/central_vocabulary.duckdb' AS central_vocab (READ_ONLY)")
cohort = "derived/ederri_features/ederri_features_Temple.parquet"

res = extract_sparse_concept_matrix(con, cohort, min_patient_freq=50)
res["X"].shape            # (N, V) csr_matrix; res["concepts"] has names, domains, patient counts

sard = extract_sard_visit_tensors(con, cohort, min_patient_freq=50)
sard["concept_tensor"].shape   # (N, max_nvisits, max_visit_len); also visit_lengths, n_visits, times

library(DBI); library(duckdb)
con <- dbConnect(duckdb::duckdb(), "derived/omop_duckdb/omop_Temple.duckdb", read_only = TRUE)
dbExecute(con, "ATTACH 'derived/omop_duckdb/central_vocabulary.duckdb' AS central_vocab (READ_ONLY);")
res <- extract_sparse_concept_matrix(con, "derived/ederri_features/ederri_features_Temple.parquet",
                                     min_patient_freq = 50)
dim(res$X)

Conventions follow omop-learn so outputs drop into its models: special tokens [BOS]=0 [EOS]=1 [SEP]=2 [PAD]=3 [UNK]=4, concept tokens from 5 ("<concept_id> - <domain> - <concept_name>", sorted like omop-learn’s ConceptTokenizer), visit times as days since 1900-01-01 with -1 padding. Pass a train result’s tokenizer (R: tokens) to a test cohort to get identical columns; pass explicit max_nvisits / max_visit_len for a fixed tensor shape. pip install "omop-duck-db[ml]" adds scipy.

2. omop-learn backend

omop-learn ships Postgres, BigQuery and Spark backends behind one interface; DuckDBBackend adds DuckDB. omop-learn’s per-patient feature SQL runs as a single LATERAL join instead of one query per patient (about 70x faster in our benchmark), with DuckDB-dialect feature SQL bundled in inst/sql/omop_learn.

from pathlib import Path
from omop_learn.omop import OMOPDataset
from omop_learn.utils.config import Config
from omop_etl import DuckDBBackend, default_features, cohort_from_parquet

backend = DuckDBBackend(con=con)
dataset = OMOPDataset(name="temple", config=Config({"cdm_schema": "main"}),
                      cohort=cohort_from_parquet(con, cohort), features=default_features(),
                      backend=backend, data_dir=Path("datasets"))
torch_ds = dataset.to_torch()      # omop-learn's own datasets and models from here on

omop-learn is not on PyPI: pip install git+https://github.com/clinicalml/omop-learn.

3. OHDSI PatientLevelPrediction and DeepPatientLevelPrediction

PLP, DeepPLP and FeatureExtraction read the CDM through DatabaseConnector, which opens the DuckDB file itself and cannot ATTACH a central vocabulary. The vocabulary tables must therefore be inside the file. It also opens the file read-write, so no other process (for example a Python session that still has the database open, even read-only) may hold it. hades_preflight() checks both, from a fresh connection:

hades_preflight("omop_Temple.duckdb")
db <- plp_database_details("omop_Temple.duckdb", target_id = 1, outcome_ids = 2)

# Native OHDSI features (including DeepPLP's temporal transformer input):
plp <- PatientLevelPrediction::getPlpData(db, FeatureExtraction::createCovariateSettings(
         useConditionOccurrenceLongTerm = TRUE, longTermStartDays = -365, endDays = -1),
       PatientLevelPrediction::createRestrictPlpDataSettings())

# ...or hand PLP / DeepPLP the very same matrix you used in Python or glmnet:
res <- extract_sparse_concept_matrix(con, cohort_parquet, value = "binary")
plp <- as_plp_data(res, con, target_id = 1, outcome_id = 2)

as_plp_data() uses covariate ids concept_id * 1000 + analysis_id with FeatureExtraction’s long-term analysis ids (102 / 302 / 502), so they match natively extracted covariates for the 365-day window (verified against FeatureExtraction in the test suite).

4. tidymodels and in-database scoring (orbital, tidypredict)

as_tidymodels_data() bridges sparse concept matrices into a tibble ready for tidymodels (rsample, recipes, parsnip, workflows, yardstick), with outcome y formatted with the event as the first factor level (event_level = "first"), and concept predictors named <domain>_<concept_id>.

Once a model or workflow is fitted, model_concepts() identifies which concepts have non-zero coefficients. materialize_concept_features() exports matching wide feature tables inside DuckDB using identical windowing logic. Scoring can then run directly in-database with orbital::orbital() or tidypredict without feature drift:

library(tidymodels)

res <- extract_sparse_concept_matrix(con, "cohort.parquet", min_patient_freq = 50, value = "binary")
dat <- as_tidymodels_data(res)

split <- rsample::group_initial_split(dat, group = person_id)
rec <- recipes::recipe(y ~ ., data = rsample::training(split)) |>
  recipes::update_role(person_id, index_date, new_role = "id")
spec <- parsnip::logistic_reg(penalty = 0.01, mixture = 1) |> parsnip::set_engine("glmnet")
fit <- workflows::workflow(rec, spec) |> parsnip::fit(rsample::training(split))

# In-database scoring on new patients without feature drift:
used <- model_concepts(fit, res$concepts)
materialize_concept_features(con, "new_cohort.parquet", used, table = "to_score", value = "binary")
orb <- orbital::orbital(fit)
scores <- orbital::augment(orb, dplyr::tbl(con, "to_score")) |> dplyr::collect()

Wide tables, messy timestamps, medication classes, phenotypes & resource limits (Python and R)

import duckdb
from omop_etl import (configure_resources, load_wide_measurement, vectorized_strptime,
                      extract_medication_classes, audit_medication_strategies,
                      ConditionCriteria, MeasurementCriteria, DrugCriteria, execute_phenotype,
                      DEFAULT_VITAL_CONCEPT_MAPPING, DEFAULT_VITAL_PLAUSIBILITY_FILTERS)

con = duckdb.connect("omop_cdm.duckdb")
configure_resources(con)                       # 75% of available RAM, best spill volume, cores - 1

# staged, one-concept-per-commit unpivot of a wide table, with clinical plausibility bounds
load_wide_measurement(con, "vital.csv", DEFAULT_VITAL_CONCEPT_MAPPING,
                      value_plausibility_filters=DEFAULT_VITAL_PLAUSIBILITY_FILTERS, missing_columns="skip")

# vectorized parsing of '05JAN2020' + '4:15', '01/05/2020' + '4:15 PM', ...
con.sql(f"SELECT {vectorized_strptime('MEASURE_DATE', 'MEASURE_TIME', on_error='default_date')} FROM vital_raw")

# two independent medication-class strategies and a per-site agreement audit
flags, audit = extract_medication_classes(con, "my_cohort", {"statin": {"atc_codes": ["C10AA"]}},
                                          strategy="both", crosswalk="site_drug_crosswalk.csv", return_audit=True)
print(audit[audit.flagged])                    # kappa, concordance, reason per site/class

# composable phenotype, compiled to one query
t2dm = (ConditionCriteria(icd_codes=["E11"], concept_ids=[201826])
        | MeasurementCriteria(loinc_codes=["4548-4"], value_gte=6.5)
        | DrugCriteria(concept_ids=[21600712], min_days_supply=30))
execute_phenotype(con, t2dm & ~ConditionCriteria(icd_codes=["E10"]), "cohort", cohort_definition_id=7)

The same functions exist in R with identical semantics (snake_case, library(omopduckdb)):

library(omopduckdb)
con <- DBI::dbConnect(duckdb::duckdb(), "omop_cdm.duckdb")
configure_resources(con)
load_wide_measurement(con, "vital.csv", DEFAULT_VITAL_CONCEPT_MAPPING,
                      value_plausibility_filters = DEFAULT_VITAL_PLAUSIBILITY_FILTERS, missing_columns = "skip")
DBI::dbGetQuery(con, paste("SELECT", vectorized_strptime("MEASURE_DATE", "MEASURE_TIME", on_error = "default_date"), "FROM vital_raw"))
res <- extract_medication_classes(con, "my_cohort", list(statin = list(atc_codes = "C10AA")),
                                  strategy = "both", crosswalk = "site_drug_crosswalk.csv", return_audit = TRUE)
res$audit[res$audit$flagged, ]
t2dm <- condition_criteria(icd_codes = "E11", concept_ids = 201826) |
  measurement_criteria(loinc_codes = "4548-4", value_gte = 6.5) |
  drug_criteria(concept_ids = 21600712, min_days_supply = 30)
execute_phenotype(con, t2dm & !condition_criteria(icd_codes = "E10"), "cohort", cohort_definition_id = 7)

All file paths handed to DuckDB go through omop_etl.paths.sanitize_sql_path, so Windows paths work in ATTACH, read_csv, read_parquet and COPY.

Clinical Risk Scores, Anthropometrics & Consortium Federation (Python & R)

1. Bedside Benchmark Heuristics (LACE Index & HOSPITAL Score)

Evaluate standardized clinical heuristics for 30-day readmissions and mortality directly inside DuckDB:

from omop_duck_db.scores import calculate_lace_index, calculate_hospital_score

# Compute LACE (Length of stay, Acuity, Charlson, ED visits)
lace_df = calculate_lace_index(con, cohort_table="inpatient_cohort")

# Compute HOSPITAL score (Hemoglobin, Oncology, Sodium, Procedure, Index type, Admissions, LOS)
hosp_df = calculate_hospital_score(con, cohort_table="inpatient_cohort")

In R:

lace_df <- calculate_lace_index(con, cohort_table = "inpatient_cohort")
hosp_df <- calculate_hospital_score(con, cohort_table = "inpatient_cohort")

2. Anthropometric Harmonization & Derived BMI

Standardize inches ([in_us]) and pounds ([lb_av]) to metric units (cm and kg), recalculate BMI $= \text{weight} / \text{height}^2$ , and replace implausible raw values ( $< 10$ or $> 120$ ):

from omop_etl import harmonize_anthropometrics

report = harmonize_anthropometrics(con, create_view=True, replace_implausible_bmi=True)
# Created view: measurement_harmonized_anthropometrics

In R:

report <- harmonize_anthropometrics(con, create_view = TRUE, replace_implausible_bmi = TRUE)

3. Transitions-of-Care Medication Trajectories

Track medication continuation, initiation, and discontinuation across care transitions:

from omop_etl import extract_medication_transitions

transitions = extract_medication_transitions(con, cohort_table="inpatient_cohort")
trans_df = transitions.to_dataframe()
# Generates {class}_initiated_inpatient, {class}_discontinued_at_discharge, {class}_continued_across_stay
summary = transitions.summary()
ct = transitions.transition_matrix("raas_inhibitors")

In R:

trans <- extract_medication_transitions(con, cohort_table = "inpatient_cohort")
df <- trans$to_dataframe()
summ <- trans$summary()
matrix <- trans$transition_matrix("raas_inhibitors")

4. Multi-Site Federated Consortium Connector

Orchestrate multi-center studies (e.g. eDERRI) with independent site databases attached to a shared Athena vocabulary:

from omop_duck_db import FederatedConsortium

consortium = FederatedConsortium(
    sites={
        "Site A": "omop_Temple.duckdb",
        "Site B": "omop_PSU.duckdb",
        "Site C": "omop_Hopkins.duckdb"
    },
    central_vocabulary="central_vocabulary.duckdb"
)

# Cross-site Table 1 and Table 2 in a single call:
t1 = consortium.generate_table1(domain="heart_failure")
t2 = consortium.generate_table2(cohort_table="v_visit_occurrence")
audit = consortium.site_summaries()

# 5. Consortium Bedside Scoring
scores = consortium.calculate_scores(cohort_table="inpatient_cohort")
print(scores.summary())
consortium.close()

In R:

fc <- federated_consortium(
  sites = c("Site A" = "omop_Temple.duckdb", "Site B" = "omop_PSU.duckdb"),
  central_vocabulary = "central_vocabulary.duckdb"
)
t1 <- consortium_generate_table1(fc, domain = "heart_failure")
t2 <- consortium_generate_table2(fc, cohort_table = "v_visit_occurrence")
audit <- consortium_site_summaries(fc)
scores <- consortium_calculate_scores(fc, cohort_table = "inpatient_cohort")
summary(scores)
consortium_close(fc)

5. Cross-Site Model Calibration & Decile Drift

Quantify model transportability and risk calibration drift across consortium sites via Greenwood-Nam-D’Agostino $\chi^2$ goodness-of-fit, Brier score decomposition, and decile tracking:

from omop_duck_db.ml import consortium_calibration

cal = consortium_calibration(
    y_true=y_val,
    y_prob=oof_probs,
    site_labels=site_series,
    n_bins=10
)
print(cal.summary())
fig = cal.plot()

In R:

cal <- consortium_calibration(
  y_true = y_val,
  y_prob = oof_probs,
  site_labels = site_series,
  n_bins = 10
)
summary(cal)
plot(cal)

6. Clinical Decision Curve Analysis (DCA)

Demonstrate clinical utility of risk stratification models across actionable decision thresholds compared to “treat all” and “treat none” defaults:

from omop_duck_db.ml import decision_curve_analysis

dca = decision_curve_analysis(
    y_true=y_val,
    y_prob=oof_probs,
    threshold_range=(0.05, 0.35)
)
print(dca.summary())
fig = dca.plot()

In R:

dca <- decision_curve_analysis(
  y_true = y_val,
  y_prob = oof_probs,
  threshold_range = c(0.05, 0.35)
)
summary(dca)
plot(dca)

7. Competing Risks Cumulative Incidence Function (CIF) & Fine-Gray Modeling

Estimate non-parametric Aalen-Johansen CIF curves and Fine-Gray subdistribution hazards for 30-day readmissions accounting for inpatient and out-of-hospital mortality:

from omop_duck_db.survival import estimate_cif, fine_gray_estimator, plot_cif

cif = estimate_cif(surv_df, time_col="time_days", status_col="status")
print(cif.summary())
fig = plot_cif(cif)

fg = fine_gray_estimator(surv_df, time_col="time_days", status_col="status", covar_cols=["age", "charlson_score"])
print(fg.summary())

In R:

cif <- estimate_cif(surv_df, time_col = "time_days", status_col = "status")
summary(cif)
plot_cif(cif)

fg <- fine_gray_estimator(surv_df, time_col = "time_days", status_col = "status", covar_cols = c("age", "charlson_score"))
summary(fg)

Testing

tests/fixtures/pcornet_sample/ is a small set of fabricated (not real) sample rows used to smoke-test both ETLs and check they produce identical output.

R: devtools::test() (or R CMD check) runs the testthat suite.

Python: pytest runs the Python-side tests, including a cross-language parity check that runs both ETLs against the fixture and diffs every row (skipped automatically if R isn’t available on the machine running the tests).

The ML bridge tests skip what they cannot import: omop-learn (point OMOP_LEARN_SRC at a source checkout’s src/ directory, or install it) and the HADES packages (FeatureExtraction, PatientLevelPrediction, Andromeda, DatabaseConnector).

Large files (Google Drive)

Instead of git/git-LFS, the generated database and vocabulary archive can be synced through Google Drive:

gdrive_upload_snapshot()    # push the current database + vocabulary archive
gdrive_download_latest()    # pull the latest snapshot instead of rebuilding

Configure the target Drive folder in config/gdrive.yml (copy from config/gdrive.yml.example).

Known limitations

Athena vocabulary downloads do not include CPT4 concept names (AMA copyright) — a separate licensed tool (bundled with the download) populates them if you have a UMLS license. Until that’s run, CPT4-coded procedures map to concept_id 0.
A handful of source fields (e.g. admission source) aren’t mapped yet; see inline comments in the ETL code for specifics.
The Python side doesn’t yet have its own BigQuery connector — use the R connector for that, either way, into the same database file.

omop-duck-db

omop-duck-db

Scope

Layout

Installation

Developer Note: renv Workflow

Example

Building the database

Refreshing the vocabulary

Transparent Central Vocabulary & Single-Store Architecture (omop_connect)

Custom mapping and retroactive remapping

Study Cohorts, Phenotyping & Clinical Pipeline

1. Unified Study Cohort Engine (define_study_cohort)

2. Comorbidity Profiling: 31 Elixhauser Domains & Charlson Index

3. Physiologic Measurement Sanitization & Outlier Winsorization (sanitize_measurements)

4. Build Inpatient Readmission Cohort (build_readmission_cohort)

5. Extract Temporal Utilization & Demographic Features

6. Athena Concept Hierarchy Rollup & Consolidated Measurements

7. Stratified Table 1 Generation & Completeness Matrix

8. Benchmark Reconciliation & Validation Harness

ML bridge: omop-learn, scikit-learn, SARD & OHDSI PLP / DeepPLP

1. Sparse concept matrix and SARD visit tensors

2. omop-learn backend

3. OHDSI PatientLevelPrediction and DeepPatientLevelPrediction

4. tidymodels and in-database scoring (orbital, tidypredict)

Wide tables, messy timestamps, medication classes, phenotypes & resource limits (Python and R)

Clinical Risk Scores, Anthropometrics & Consortium Federation (Python & R)

1. Bedside Benchmark Heuristics (LACE Index & HOSPITAL Score)

2. Anthropometric Harmonization & Derived BMI

3. Transitions-of-Care Medication Trajectories

4. Multi-Site Federated Consortium Connector

5. Cross-Site Model Calibration & Decile Drift

6. Clinical Decision Curve Analysis (DCA)

7. Competing Risks Cumulative Incidence Function (CIF) & Fine-Gray Modeling

Testing

Large files (Google Drive)

Known limitations

Developer Note: `renv` Workflow

Transparent Central Vocabulary & Single-Store Architecture (`omop_connect`)

1. Unified Study Cohort Engine (`define_study_cohort`)

3. Physiologic Measurement Sanitization & Outlier Winsorization (`sanitize_measurements`)

4. Build Inpatient Readmission Cohort (`build_readmission_cohort`)