VKD/Start a conversation
Back to projectsINTERACTIVE INTEGRITY DEMO / SYNTHETIC CLINICAL DATA

CLINICAL TRIAL DRIFT GUARD

Find site-level bias
before analysis begins.

Inspect cross-site population stability, calibration shifts and de-identification controls.

View executable source
Detection threshold

2 of 5 sites flagged

FEATURE / SITE
Systolic BP
BMI
Age
SELECTED DIAGNOSTIC

SITE-04

BP PSI1.68
KS statistic0.778
Sensor shift+22.5 mmHg
Baseline108–132
SITE-04129–155

Escalate: distribution shift exceeds the selected PSI threshold.

MEASURED PIPELINE RESULTS

Integrity controls with quantified behavior.

10,000+Synthetic records across five sites
<2.5 secDocumented KS + PSI detection latency
0 exposed PIIDirect identifiers blocked before Parquet
Trial-site exportsMulti-site clinical records
→
PySpark privacy layerPII removal + salted hashing
→
KS + PSI engineBaseline comparison and flags
→
Quality dashboardDiagnostics before analysis

AUTOMATED VALIDATION

Pipeline assertions

  • Raw dataset generation succeeds
  • Processed Parquet artifact is created
  • Name, SSN and patient ID columns are absent
  • Deterministic masked patient ID is retained

PRODUCTION TRADEOFFS

Decisions at scale

Synthetic → governed clinical data

The demo avoids PHI completely; production needs IRB-approved access, schema contracts, lineage and controlled research workspaces.

Fixed baseline → rolling reference

A stable control improves interpretability; long trials require versioned windows and safeguards against contaminating the baseline.

Statistical flags → operational workflow

KS and PSI identify distribution change, not clinical causality. Production alerts need ownership, adjudication and documented resolution.