v3.0

RCT Evidence Schema

A normalized, versioned, auditable evidence graph for clinical trial data extraction.

14 Entity Data Model

This schema defines a normalized data model for RCT evidence extraction. Each entity has explicit relationships and every field carries provenance metadata.

Core Entities

1 Study
2 Publication
3 Estimand
4 Arm
5 Outcome
6 Endpoint
7 Analysis Population

Result & Assessment Entities

8 Result Instance
9 Comparison
10 Appraisal Evidence
11 Appraisal Judgment
12 Evidence Context
13 Forensic Flags
14 Lifecycle
15 Causal Audit

Provenance Types

reportedDirectly from paper

calculatedDerived from reported data

inferredHuman/model judgment

externalFrom external sources

synthesisRequires multiple studies

Relationships

1:N → Publication1:N → Arm1:N → Outcome1:1 → Estimand

Identity

study_id

Why capture this?

Links all study components. Essential for cross-referencing.

required: stringcalculated

Unique identifier (auto-generated)

registration_id

Why capture this?

Enables protocol-to-publication comparison. Detects outcome switching.

required: stringreported

ClinicalTrials.gov, ISRCTN, etc.

e.g., NCT01234567

acronym

Why capture this?

Common in literature. Helps readers recognize landmark trials.

: stringreported

Trial acronym if any

e.g., JUPITER

official_title

Why capture this?

Full title often contains PICO elements not in acronym.

required: stringreported

Design

design_type

Why capture this?

Determines appropriate analysis methods and bias considerations.

required: enumreported

parallel

Groups receive different treatments simultaneously. Most common RCT design.

crossover

Same participants receive both treatments in sequence. Each person serves as own control.

factorial

Tests multiple interventions simultaneously (e.g., 2×2 design). Efficient for interactions.

cluster

Randomizes groups (clinics, schools) not individuals. Requires ICC adjustment.

adaptive

Design changes based on interim results. Sample size, allocation, or arms may adapt.

platform

Master protocol testing multiple treatments. Arms can be added/dropped over time.

phase

Why capture this?

Phase affects generalizability. Phase III has different populations than Phase IV.

: enumreported

First-in-human. Safety, dosing, pharmacokinetics. Usually 20-100 healthy volunteers.

Efficacy signal. Dose-response. 100-300 patients with condition.

III

Confirmatory efficacy. 1000-3000+ patients. Basis for regulatory approval.

Post-marketing. Real-world effectiveness, rare adverse events, new populations.

N/A

Not applicable (e.g., behavioral, surgical, or non-drug interventions).

masking

Why capture this?

Critical for RoB assessment. Performance and detection bias depend on masking.

: enumreported

open-label

No blinding. Everyone knows assignment. High bias risk for subjective outcomes.

single-blind

Usually participants blinded, investigators not. Reduces placebo effect.

double-blind

Participants and investigators blinded. Gold standard for bias reduction.

triple-blind

Participants, investigators, and outcome assessors/analysts blinded.

allocation

Why capture this?

Fundamental to causal inference validity.

: enumreported

randomized

Random assignment to groups. Balances known and unknown confounders.

non-randomized

Assignment by other mechanism (preference, availability). Confounding likely.

Sites & Geography

centers

Why capture this?

Multicenter reduces site-specific effects but adds coordination challenges.

: enumreported

single-center

One site. May have site-specific effects. Limited generalizability.

multicenter

Multiple sites. Better generalizability but coordination complexity.

n_sites

Why capture this?

Affects generalizability and heterogeneity assessment.

: numberreported

countries

Why capture this?

Geographic context affects applicability to specific populations.

: string[]reported

e.g., ['USA', 'Canada', 'UK']

Timeline

recruitment_start

Why capture this?

Historical context matters. Pre-COVID trials may not apply post-COVID.

: datereported

recruitment_end

Why capture this?

Slow recruitment may indicate protocol issues or population mismatch.

: datereported

study_completion

Why capture this?

Time from completion to publication can indicate reporting issues.

: datereported

total_duration_months

Why capture this?

Duration affects power for time-to-event outcomes and adverse event detection.

: numbercalculated

Funding & Oversight

funding_source

Why capture this?

Industry funding associated with positive results bias. Critical for COI assessment.

: enumreported

industry

Pharmaceutical, device, or other commercial company funding.

public

Government grants (NIH, MRC, etc.) or nonprofit foundations.

mixed

Both industry and public/nonprofit funding.

none

No external funding reported (investigator-initiated).

undisclosed

Funding source not stated. Red flag for transparency.

sponsor

Why capture this?

Specific sponsors have track records. Enables pattern detection.

: stringreported

sponsor_role

Why capture this?

'Sponsor designed and analyzed' vs 'independent statistician' matters for bias.

: textreported

Involvement in design, conduct, analysis

ethics_approval

Why capture this?

Missing ethics approval is a red flag for study integrity.

: booleanreported

ethics_body

Why capture this?

Some ethics bodies are more rigorous. Enables pattern detection.

: stringreported

dsmb

Why capture this?

DSMB oversight reduces risk of inappropriate early stopping.

: booleanreported

Data Safety Monitoring Board

Protocol

protocol_doi

Why capture this?

Protocol enables outcome switching detection. Published protocols are more trustworthy.

: stringreported

protocol_version

Why capture this?

Multiple versions may indicate protocol amendments worth examining.

: stringreported

sap_available

Why capture this?

Pre-specified SAP prevents analytical flexibility abuse.

: booleanreported

Statistical Analysis Plan

Relationships

1:1 → Studyreferenced by → Comparison

Why Estimand matters: Without explicit estimand, we extract numbers but miss what is actually being estimated. The same trial can target different estimands (ITT vs per-protocol, treatment policy vs hypothetical).

ICH E9 R1 Components

estimand_id

Why capture this?

Links estimand to comparisons. A study can have multiple estimands.

required: stringcalculated

population

Why capture this?

The causal effect is specific to a population. Different from study eligibility criteria.

required: textreported

Target population for the causal question

e.g., Adults ≥40 with LDL ≥130 mg/dL and no prior CVD

treatment

Why capture this?

Full specification including dose, route, duration. Not just drug name.

required: textreported

Treatment condition

e.g., Atorvastatin 40mg daily for 24 months

comparator

Why capture this?

Placebo vs active comparator matters for interpretation. Usual care is context-dependent.

required: textreported

Comparator condition

e.g., Matching placebo daily for 24 months

outcome_id

Why capture this?

Links to outcome definition. Same outcome can appear in multiple estimands.

required: stringcalculated

Foreign key to Outcome

time_horizon

Why capture this?

Short-term benefit may not equal long-term benefit. Essential for interpretation.

required: stringreported

e.g., 24 weeks, 5 years, median follow-up 4.2 years

summary_measure

Why capture this?

RR vs OR vs HR are not interchangeable. Choice affects interpretation.

required: enumreported

mean_difference

Difference in means. For continuous outcomes. Units matter.

risk_ratio

Ratio of risks (RR). P(event|treatment) / P(event|control). Multiplicative.

odds_ratio

Ratio of odds (OR). Approximates RR when events rare. Common in logistic regression.

hazard_ratio

Ratio of instantaneous event rates (HR). For time-to-event outcomes.

risk_difference

Absolute risk reduction (ARR). P(event|control) - P(event|treatment).

rate_ratio

Ratio of incidence rates (IRR). Events per person-time.

Intercurrent Event Handling

ic_events

Why capture this?

Treatment discontinuation, rescue medication, death — how these are handled defines what you're estimating.

: object[]reported

How intercurrent events are handled

Intercurrent Event Strategy

<intercurrent_event>
  <event>treatment_discontinuation</event>
  <strategy>treatment_policy</strategy>
  <description>Analyze all participants as randomized regardless of adherence</description>
</intercurrent_event>

<intercurrent_event>
  <event>rescue_medication</event>
  <strategy>composite</strategy>
  <description>Rescue medication use counted as treatment failure</description>
</intercurrent_event>

<!-- Strategy options: treatment_policy | hypothetical | composite | principal_stratum | while_on_treatment -->

Entity Relationships


┌─────────────────────────────────────────────────────────────────┐
│                           STUDY                                 │
└─────────────────────────────────────────────────────────────────┘
                              │
       ┌──────────────────────┼──────────────────────┐
       │                      │                      │
       ▼                      ▼                      ▼
┌─────────────┐        ┌─────────────┐        ┌─────────────┐
│ PUBLICATION │        │     ARM     │        │   OUTCOME   │
└─────────────┘        └──────┬──────┘        └──────┬──────┘
                              │                      │
                              │              ┌───────┴──────┐
                              │              ▼              │
                              │       ┌─────────────┐       │
                              │       │  ENDPOINT   │       │
                              │       │ (outcome ×  │       │
                              │       │  timepoint) │       │
                              │       └──────┬──────┘       │
                              │              │              │
                              └──────┬───────┘              │
                                     │                      │
                                     ▼                      │
                    ┌──────────────────────────┐            │
                    │    ANALYSIS POPULATION   │            │
                    └────────────┬─────────────┘            │
                                 │                          │
          ┌──────────────────────┼──────────────────────┐   │
          │                      │                      │   │
          ▼                      ▼                      ▼   │
   ┌─────────────────────────────────────────────────────┐  │
   │               RESULT INSTANCE                       │  │
   │  (arm × endpoint × population = composite key)      │  │
   └───────────────────────┬─────────────────────────────┘  │
                           │                                │
                           ▼                                │
   ┌─────────────────────────────────────────────────────┐  │
   │              COMPARISON                             │  │
   │  (result₁ vs result₂, bound to estimand)            │◄─┘
   └───────────────────────┬─────────────────────────────┘
                           │                              ┌─────────────┐
                           │                              │  ESTIMAND   │
          ┌────────────────┼────────────────┬─────────┐   │ (ICH E9 R1) │
          │                │                │         │   └─────────────┘
          ▼                ▼                ▼         ▼          ▲
   ┌────────────┐   ┌────────────┐   ┌────────────┐  ┌───────────────────┐
   │ APPRAISAL  │   │ APPRAISAL  │   │ EVIDENCE   │  │   CAUSAL AUDIT    │
   │ EVIDENCE   │   │ JUDGMENT   │   │ CONTEXT    │  │ (bridges RCT→DAG) │
   └─────┬──────┘   └─────┬──────┘   └────────────┘  │ • identification  │
         │                │                          │ • adjustment set  │
         └────────┬───────┘                          │ • threats & flags │
                  │                                  └───────────────────┘
   ┌──────────────┼──────────────┐
   │                             │
   ▼                             ▼
┌────────────────┐     ┌─────────────────┐
│ FORENSIC FLAGS │     │    LIFECYCLE    │
│ (optional,     │     │ (field-level,   │
│  high-variance)│     │  actor+conf+ver)│
└────────────────┘     └─────────────────┘

RCT Evidence Schema

14 Entity Data Model

Core Entities

Result & Assessment Entities

Provenance Types

🔬Study

Identity

Design

Sites & Geography

Timeline

Funding & Oversight

Protocol

📄Publication

🎯Estimand

ICH E9 R1 Components

Intercurrent Event Handling

💊Arm

📊Outcome

⏱️Endpoint

👥Analysis Population

📈Result Instance

⚖️Comparison

🔍Appraisal Evidence

⚠️Appraisal Judgment

🌐Evidence Context

🔎Forensic Flags

📜Lifecycle

🎯Causal Audit

Entity Relationships