AI Ethics Data Provenance and Consent Compliance Dashboard

A Synthetic Framework for Visualizing Ethical Risk and Data Governance in Clinical AI Systems

1. Introduction

The proliferation of Artificial Intelligence (AI) in healthcare—spanning clinical decision support systems (CDSS), electronic health record (EHR) analytics, and medical imaging diagnostics—has introduced profound challenges in data ethics, consent management, and provenance traceability.

As healthcare datasets expand through federated learning architectures, data provenance (the ability to trace the origin, consent, and usage of data elements) becomes pivotal for regulatory compliance under frameworks such as GDPR (General Data Protection Regulation), HIPAA (Health Insurance Portability and Accountability Act), PIPEDA (Personal Information Protection and Electronic Documents Act, Canada), and TCPS2 (Tri-Council Policy Statement: Ethical Conduct for Research Involving Humans).

This project—the AI Ethics Data Provenance Explorer—was designed to provide a synthetic yet realistic simulation of how patient consent, dataset lineage, and AI model usage can be encoded, analyzed, and visualized interactively. It integrates:

• Synthetic data generation (via Faker and NumPy) to emulate real-world clinical and consent variability.
• Ethical compliance computation using a formalized scoring engine.
• Visualization through D3.js for interactive exploration of bias, scope violations, and provenance flows.

This system demonstrates how ethics-aware analytics can be embedded directly into the data lifecycle—from acquisition to AI model deployment.

2. Data Architecture and Schema Design

The synthetic data engine generates multi-table relational structures reflecting the intersection between clinical and ethical domains.

Each record in the merged dataset represents a specific AI model’s interaction with a patient record under a given consent context.

2.1 Data Generation Logic

Randomized, but biologically and statistically coherent, distributions were created:

ages = rng.integers(18, 90, size=n)
sexes = rng.choice(["Female", "Male", "Other"], size=n, p=[0.49, 0.49, 0.02])
diseases = rng.choice(["Diabetes", "Hypertension", "Cancer", "COPD", "Sepsis"],
                      size=n, p=[0.27, 0.28, 0.22, 0.13, 0.10])

The consent schema introduces ethical variability:

• Specific consent: restricted to one project or purpose.
• Broad consent: reusable for related AI research.
• None/withdrawn: ethically invalidated.
• Expired: time-bounded consent period exceeded.

Temporal logic ensures heterogeneity in consent expiration:

expired_flags = np.array([ce.date() < today for ce in consent_end], dtype=bool)

This approach mirrors the real-world entropy of institutional review board (IRB) records, where overlapping consent states complicate AI data reuse.

3. Ethical Compliance Engine

3.1 Scoring Model

A deterministic Compliance Scoring Function (CSF) computes a 0–100 ethical compliance score per record, applying additive penalties for violations.

Core algorithmic block:

def compute_compliance_score(row):
    score = 100
    violations = []

    if row["withdrawn"]:
        score -= 100; violations.append("withdrawn_use")

    if used > consent_end:
        score -= 80; violations.append("expired_use")

    if row["consent_type"] == "none":
        score -= 80; violations.append("no_consent_use")

    if row["consent_type"] == "specific" and row["purpose"] != row["intended_purpose"]:
        score -= 40; violations.append("consent_scope_mismatch")

    if row["consent_type"] == "broad" and row["intended_purpose"] == "commercial_ai":
        score -= 25; violations.append("commercial_overreach")

    if pd.isna(row["source_dataset"]) or pd.isna(row["hospital_id"]):
        score -= 20; violations.append("provenance_gap")

    return max(0, min(100, score)), violations[0] if violations else "none"

3.2 Risk Stratification

Compliance scores are binned:

• Compliant: ≥70
• At Risk: 40–69
• Non-Compliant: <40

This enables downstream ethical stratification analytics, akin to clinical risk scoring.

4. Data Visualization Framework (D3.js)

Visualization translates compliance metadata into interactive intelligence. The dashboard leverages D3.js v7 and d3-sankey to encode relationships across hierarchical ethical layers.

4.1 Visualization Stack

5. Implementation Highlights

5.1 Data Binding and Animation

D3 transitions were used to provide a cognitive affordance for data change and narrative continuity:

paths.transition()
     .duration(1000)
     .attrTween("d", function(d) {
        const i = d3.interpolate({ startAngle: 0, endAngle: 0 }, d);
        return t => arc(i(t));
     });

This yields a radial interpolation from 0 to full arc extent, animating consent categories into view.

5.2 Dynamic Scales and Palettes

Instead of D3’s default categorical schemes, a custom pastel palette was pre-computed in Python and serialized into JSON:

{
  "specific": "#A3D5FF",
  "broad": "#FFC9DE",
  "none": "#CDECCF",
  "withdrawn": "#FFE3A3"
}

Ensures visual harmony and accessibility under low-contrast constraints, aligning with ISO 9241-210 (Human-Centered Design Ergonomics).

5.3 Sankey Construction

Hierarchical aggregation constructs multi-level provenance flows:

for _, r in df.groupby(["model_name","reuse_category"]).size().reset_index(name="count").iterrows():
    links.append({"source": add_node(r["model_name"]),
                  "target": add_node(r["reuse_category"]),
                  "value": int(r["count"])})

Resultant Sankey nodes capture the semantic lineage of data reuse, rendering visible the cascade from consent origin to ultimate AI model exploitation.

6. Analytical Outputs

6.1 Quantitative Compliance Summaries

The system auto-generates analytical summaries for:

• Hospital Compliance Metrics (compliance_summary_hospital.csv)
• Model Compliance Metrics (compliance_summary_model.csv)

Each includes: mean_score, pct_at_risk, pct_non_compliant, n.

These datasets support meta-level ethics benchmarking, enabling institutional dashboards or regulatory audits.

6.2 JSON Data for Web Integration

All computed artifacts are encapsulated in a single D3-ready JSON (ethics_dashboard_data.json):

{
  "donutConsent": [...],
  "statusByHospital": {...},
  "timelineActiveExpired": {...},
  "heatModelHospital": {...},
  "sankey": {...}
}

This structure allows asynchronous ingestion in client-side rendering frameworks such as SvelteKit, React, or Vue without backend dependencies.

7. Deployment Architecture

7.1 Data Flow

Google Colab → Google Drive (/............./data/)
             ↓
Firebase Hosting (static/data/json/)
             ↓
SvelteKit Front-End (ethics-dashboard.svx)
             ↓
Client-Side D3 Rendering

7.2 Tech Stack Summary

8. Usefulness and Applications

8.1 Educational Sandbox

This project functions as a pedagogical model for universities and research ethics boards to demonstrate consent drift and data overreach scenarios without using any real PHI (Protected Health Information).

8.2 Regulatory Simulation

Agencies can simulate GDPR Article 30 “Records of Processing Activities (RoPA)” visualizations—auditing lineage, retention, and lawful basis for AI-driven healthcare analytics.

8.3 AI Governance Integration

The compliance metrics can feed into broader AI governance pipelines, e.g.:

• MLflow or Vertex AI Model Registry for model metadata tagging.
• Integration with FAIR (Findable, Accessible, Interoperable, Reusable) metadata ontologies.
• Risk scoring input for NIST AI RMF (Risk Management Framework) dashboards.

8.4 Bias & Fairness Correlation

Because demographic data (age, sex, disease type) coexists with compliance metrics, one can run secondary analytics:

corr = df.groupby("sex")["compliance_score"].mean()

Yielding insights into whether specific patient subgroups disproportionately experience ethical violations—an intersection between AI fairness and data governance.

9. Visualization Semantics

Each visualization element has explicit ethical meaning:

Combined, they create a multi-modal ethical telemetry environment, merging visual analytics with regulatory semantics.

10. Critical Code Artifacts

10.1 Compliance Score Derivation

score, tag = compute_compliance_score(row)
df["risk_band"] = pd.cut(df["compliance_score"],
                         bins=[-1,39,69,100],
                         labels=["non_compliant","at_risk","compliant"])

10.2 D3 Binding for Heatmap

svg.selectAll("rect.cell")
   .data(cells)
   .enter()
   .append("rect")
   .attr("x", d => x(d.model))
   .attr("y", d => y(d.hospital))
   .attr("width", x.bandwidth())
   .attr("height", y.bandwidth())
   .attr("fill", d => c(d.value))
   .transition()
   .duration(800)
   .style("opacity", 1);

10.3 Sankey Flow Generation

const sankey = d3Sankey().nodeWidth(16).nodePadding(12);
const { nodes, links } = sankey(graph);

These illustrate the fusion of data ethics computation and interactive visualization pipelines.

11. Limitations and Future Enhancements

Current Constraints

• Data is synthetic; no real clinical system integration.
• Consent modeling is rule-based, not probabilistic.
• Provenance granularity limited to three hops (Dataset → Model → Use-type).

Planned Extensions

1. Differential Privacy Simulation: Integrate ε-differential privacy mechanisms for anonymization sensitivity analysis.
2. Federated Provenance Mapping: Extend Sankey topology to simulate federated data nodes under HL7 FHIR (Fast Healthcare Interoperability Resources).
3. Explainable Compliance AI: Train gradient-boosted or SHAP-interpreted models to predict likely future ethical breaches.
4. Ontology Alignment: Integrate consent types with SNOMED-CT, LOINC, and OBI (Ontology for Biomedical Investigations) codes for semantic interoperability.

12. Conclusion

The AI Ethics Data Provenance Explorer represents an end-to-end simulation of ethical data governance in AI-driven healthcare pipelines.

By unifying synthetic data generation, automated ethical scoring, and interactive visualization, it operationalizes otherwise abstract bioethical principles—autonomy, beneficence, non-maleficence, and justice—into tangible analytics.

Through its modular design, pastel visualization aesthetics, and open architecture, this system can serve as a blueprint for compliance-aware AI systems, aligning technical practice with regulatory ethics and human-subject protections.