はじめに
HADES (Health Analytics Data-to-Evidence Suite) は、OHDSI の 20 以上の R パッケージのオープンソース セットで、コホート生成から推定/予測までの分析パイプライン全体を提供します。 ATLAS は GUI であり、HADES はその背後にあるエンジンです。
Quan hệ ATLAS ↔ HADES:
┌──────────────────────────────────┐
│ ATLAS (GUI) │
│ Design → Generate R Package │
│ │ │
│ ▼ │
│ ┌─────────────┐ │
│ │ HADES (R) │ │
│ │ Engine │ │
│ └─────────────┘ │
│ │ │
│ ┌─────────┴──────────┐ │
│ ▼ ▼ │
│ CohortMethod PatientLevel │
│ (Estimation) Prediction │
└──────────────────────────────────┘
→ ATLAS generate study package dùng HADES packages
→ Hoặc dùng HADES trực tiếp từ R (linh hoạt hơn)
1. HADES エコシステム
1.1 主なパッケージ
HADES Package Map:
┌─ Data Infrastructure ──────────────────────┐
│ DatabaseConnector Kết nối DB via JDBC │
│ SqlRender SQL cross-platform │
│ Eunomia CDM test database │
└────────────────────────────────────────────┘
┌─ Cohort Generation ───────────────────────┐
│ CohortGenerator Tạo & execute cohort │
│ CirceR Cohort definition → SQL │
│ CohortDiagnostics Đánh giá chất lượng │
│ cohort │
└────────────────────────────────────────────┘
┌─ Characterization ────────────────────────┐
│ FeatureExtraction Trích xuất features │
│ CohortExplorer Explore cohort data │
│ Characterization Cohort comparison │
└────────────────────────────────────────────┘
┌─ Population Analytics ────────────────────┐
│ CohortMethod Causal inference │
│ SelfControlledCase SCCS design │
│ EvidenceSynthesis Meta-analysis │
└────────────────────────────────────────────┘
┌─ Prediction ──────────────────────────────┐
│ PatientLevelPrediction ML prediction │
│ DeepPatientLevelPred. Deep learning │
│ EnsemblePatientLevel. Ensemble models │
└────────────────────────────────────────────┘
┌─ Orchestration ───────────────────────────┐
│ Strategus Pipeline orchestrator │
│ ResultModelManager Result management │
│ ShinyAppBuilder Interactive dashboards │
└────────────────────────────────────────────┘
1.2 HADES のインストール
# Cài toàn bộ HADES
install.packages("remotes")
remotes::install_github("OHDSI/Hades")
# Hoặc cài từng package
remotes::install_github("OHDSI/CohortGenerator")
remotes::install_github("OHDSI/CohortMethod")
remotes::install_github("OHDSI/PatientLevelPrediction")
remotes::install_github("OHDSI/CohortDiagnostics")
remotes::install_github("OHDSI/Strategus")
2. コホートジェネレーター
2.1 R からコホートを作成する
library(CohortGenerator)
library(DatabaseConnector)
connectionDetails <- createConnectionDetails(
dbms = "postgresql",
server = "localhost/ohdsi",
user = "ohdsi_app",
password = keyring::key_get("ohdsi"),
port = 5432
)
# Load cohort definition (JSON từ ATLAS export)
cohortDefinitionSet <- getCohortDefinitionSet(
settingsFileName = "inst/settings/CohortsToCreate.csv",
jsonFolder = "inst/cohorts",
sqlFolder = "inst/sql/sql_server"
)
# Create cohort tables
cohortTableNames <- getCohortTableNames(
cohortTable = "my_study_cohort"
)
createCohortTables(
connectionDetails = connectionDetails,
cohortDatabaseSchema = "results",
cohortTableNames = cohortTableNames
)
# Generate cohorts
cohortsGenerated <- generateCohortSet(
connectionDetails = connectionDetails,
cdmDatabaseSchema = "cdm",
cohortDatabaseSchema = "results",
cohortTableNames = cohortTableNames,
cohortDefinitionSet = cohortDefinitionSet
)
print(cohortsGenerated)
# cohortId | cohortName | status | count
# 1 | New-Onset Type 2 DM | COMPLETE | 4,200
# 2 | Acute MI | COMPLETE | 890
# 3 | Metformin New Users | COMPLETE | 3,100
3. コホート法 (推定)
3.1 研究デザイン
library(CohortMethod)
# Target-Comparator-Outcome
tcos <- createTargetComparatorOutcomes(
targetId = 1, # Metformin
comparatorId = 2, # Sulfonylurea
outcomeIds = c(3, 4), # MI, Stroke
excludedCovariateConceptIds = c()
)
targetComparatorOutcomesList <- list(tcos)
# Covariate settings
covSettings <- createDefaultCovariateSettings(
excludedCovariateConceptIds = c(),
addDescendantsToExclude = TRUE
)
# Study parameters
getDbCmDataArgs <- createGetDbCohortMethodDataArgs(
washoutPeriod = 365,
maxCohortSize = 0, # no limit
covariateSettings = covSettings
)
# Propensity Score
createPsArgs <- createCreatePsArgs(
maxCohortSizeForFitting = 150000,
control = createControl(
cvType = "auto",
startingVariance = 0.01,
tolerance = 2e-07,
noiseLevel = "quiet"
)
)
# Matching
matchOnPsArgs <- createMatchOnPsArgs(
maxRatio = 1, # 1:1 matching
caliper = 0.2,
caliperScale = "standardized logit"
)
# Outcome model
fitOutcomeModelArgs <- createFitOutcomeModelArgs(
modelType = "cox",
stratified = FALSE
)
3.2 実行と結果
# Execute analysis
result <- runCmAnalyses(
connectionDetails = connectionDetails,
cdmDatabaseSchema = "cdm",
exposureDatabaseSchema = "results",
exposureTable = "my_study_cohort",
outcomeDatabaseSchema = "results",
outcomeTable = "my_study_cohort",
outputFolder = "output/cm_results",
cmAnalysisList = list(cmAnalysis),
targetComparatorOutcomesList = targetComparatorOutcomesList
)
# Get results
analysisSummary <- summarizeAnalyses(result)
print(analysisSummary)
# target | comparator | outcome | rr | ci95lb | ci95ub | p
# Met | SU | MI | 0.62 | 0.39 | 0.98 | 0.041
# Met | SU | Stroke | 0.78 | 0.55 | 1.12 | 0.18
4. 患者レベルの予測
4.1 モデル設計
library(PatientLevelPrediction)
# Covariates
covSettings <- createCovariateSettings(
useDemographicsGender = TRUE,
useDemographicsAge = TRUE,
useConditionOccurrenceLongTerm = TRUE,
useDrugExposureLongTerm = TRUE,
useMeasurementValueLongTerm = TRUE,
useProcedureOccurrenceLongTerm = TRUE,
longTermStartDays = -365,
endDays = 0
)
# Population settings
populationSettings <- createStudyPopulationSettings(
washoutPeriod = 365,
firstExposureOnly = TRUE,
removeSubjectsWithPriorOutcome = TRUE,
riskWindowStart = 1,
riskWindowEnd = 1825, # 5 years
minTimeAtRisk = 365
)
# Model settings
lasso <- setLassoLogisticRegression()
gbm <- setGradientBoostingMachine(
ntrees = c(100, 300),
maxDepth = c(4, 6),
learnRate = c(0.01, 0.1)
)
4.2 実行と評価
# Run prediction
plpResults <- runPlp(
plpData = plpData,
outcomeId = 3, # CKD
analysisId = "ckd_prediction",
analysisName = "CKD Risk in DM",
populationSettings = populationSettings,
splitSettings = createDefaultSplitSetting(
trainFraction = 0.75,
testFraction = 0.25,
nfold = 3
),
modelSettings = lasso,
executeSettings = createExecuteSettings(
runSplitData = TRUE,
runSampleData = FALSE,
runfeatureEngineering = FALSE,
runPreprocessData = TRUE,
runModelDevelopment = TRUE,
runCovariateSummary = TRUE
)
)
# Performance
performance <- plpResults$performanceEvaluation
cat("AUC:", performance$evaluationStatistics$AUC, "\n")
cat("AUPRC:", performance$evaluationStatistics$AUPRC, "\n")
cat("Brier:", performance$evaluationStatistics$BrierScore, "\n")
# Kết quả:
# AUC: 0.82
# AUPRC: 0.35
# Brier: 0.08
4.3 光沢のある結果を表示する
# Launch interactive viewer
viewPlp(plpResults, outputFolder = "output/plp_results")
# → Mở browser:
# - ROC curve
# - Calibration plot
# - Feature importance
# - Decision curve analysis
# - Demographic parity
5. コホート診断
5.1 コホートの品質評価
library(CohortDiagnostics)
# Chạy diagnostics
executeDiagnostics(
cohortDefinitionSet = cohortDefinitionSet,
connectionDetails = connectionDetails,
cohortTable = "my_study_cohort",
cohortDatabaseSchema = "results",
cdmDatabaseSchema = "cdm",
exportFolder = "output/diagnostics",
databaseId = "Hospital_VN",
minCellCount = 5
)
# Xem kết quả
createMergedResultsFile(
dataFolder = "output/diagnostics",
sqliteDbPath = "output/MergedCohortDiagnosticsData.sqlite"
)
launchDiagnosticsExplorer(
sqliteDbPath = "output/MergedCohortDiagnosticsData.sqlite"
)
5.2 診断レポート
CohortDiagnostics cung cấp:
1. Cohort Count & Attrition
→ Bao nhiêu BN pass mỗi inclusion criteria?
2. Incidence Rate (over time)
→ Trend: tăng/giảm/ổn định?
3. Time Distribution
→ Observation time trước/sau index date
4. Concept Set Diagnostics
→ Orphan concepts (concepts bị miss)
→ Included source codes
→ Resolved concept set details
5. Index Event Breakdown
→ Source codes nào trigger cohort entry?
6. Visit Context
→ Inpatient, outpatient, ER?
7. Overlap
→ Bao nhiêu BN thuộc nhiều cohorts?
6. Strategus — パイプライン オーケストレーター
6.1 概念
Strategus = orchestrate toàn bộ analysis pipeline
Thay vì viết code riêng cho từng bước:
1. CohortGenerator → tạo cohort
2. CohortDiagnostics → kiểm tra
3. Characterization → mô tả
4. CohortMethod → estimation
5. PatientLevelPrediction → prediction
Strategus ghép tất cả thành 1 pipeline JSON:
analysisSpecifications.json → Execute once → All results
6.2 分析仕様の作成
library(Strategus)
# Cohort generation module
cohortGenModule <- CohortGeneratorModule$new()
cohortGenModuleSpecs <- cohortGenModule$createModuleSpecifications(
cohortDefinitionSet = cohortDefinitionSet
)
# Characterization module
characterizationModule <- CharacterizationModule$new()
charModuleSpecs <- characterizationModule$createModuleSpecifications(
targetIds = c(1, 2),
outcomeIds = c(3),
minPriorObservation = 365
)
# Estimation module
estimationModule <- CohortMethodModule$new()
estModuleSpecs <- estimationModule$createModuleSpecifications(
cmAnalysisList = list(cmAnalysis),
targetComparatorOutcomesList = targetComparatorOutcomesList
)
# Combine all modules
analysisSpecifications <- createEmptyAnalysisSpecificiations() |>
addModuleSpecifications(cohortGenModuleSpecs) |>
addModuleSpecifications(charModuleSpecs) |>
addModuleSpecifications(estModuleSpecs)
# Save
ParallelLogger::saveSettingsToJson(
analysisSpecifications,
"inst/analysisSpecifications.json"
)
6.3 戦略の実行
# Execution settings
executionSettings <- createCdmExecutionSettings(
connectionDetailsReference = "Hospital_VN",
workDatabaseSchema = "results",
cdmDatabaseSchema = "cdm",
cohortTableNames = getCohortTableNames("strategus_cohort"),
workFolder = "output/strategus_work",
resultsFolder = "output/strategus_results",
minCellCount = 5
)
# Store connection details (secure)
storeConnectionDetails(
connectionDetails = connectionDetails,
connectionDetailsReference = "Hospital_VN"
)
# Execute entire pipeline
execute(
analysisSpecifications = analysisSpecifications,
executionSettings = executionSettings,
executionScriptFolder = "output/execution_scripts"
)
# → Tự động chạy: CohortGen → Diagnostics → Char → Estimation
# → Kết quả lưu trong resultsFolder
7. Eunomia — CDM テスト データベース
# Eunomia: synthetic CDM database cho testing
library(Eunomia)
connectionDetails <- getEunomiaConnectionDetails()
# → Tự tạo SQLite CDM database với synthetic data
# Dùng để:
# - Test code trước khi chạy production
# - Unit testing cho R study packages
# - Training & workshops
# Ví dụ: test CohortMethod
library(CohortMethod)
cmData <- getDbCohortMethodData(
connectionDetails = connectionDetails,
cdmDatabaseSchema = "main",
targetId = 1,
comparatorId = 2,
outcomeIds = 3,
covariateSettings = createDefaultCovariateSettings()
)
summary(cmData)
# → CohortMethodData object
# → Target: 500 persons
# → Comparator: 480 persons
概要
| パッケージ | 機能 | ATLASと同等 |
|---|---|---|
| コホートジェネレーター | コホートの作成と管理 | コホートの定義 |
| コホート診断 | コホートの質の評価 | — |
| 特徴抽出 | 共変量の抽出 | 特性評価 |
| コホートメソッド | 因果推論（推定） | 見積もり |
| 患者レベル予測 | ML 予測モデル | 予測 |
| ストラテジーアス | パイプライン オーケストレーター | — |
| ユーノミア | 合成テスト CDM | — |
次の記事: OHDSI スタックを Docker と Kubernetes にデプロイする