EUCAIM Wizard Tool — ARX “project pack” for risk analysis

This repository is for data curators / analysts who want to run re-identification risk analysis and utility evaluation in the ARX Data Anonymization Tool, using the EUCAIM Wizard configuration (recommended workflow + pre-built hierarchies for DICOM metadata / clinical tabular data).

You need first:

1. export tabular data (.csv / .xlsx) containing clinical fields and/or extracted DICOM metadata (i.e. through the EUCAIM DICOM Tags Extractor),
2. load it into ARX,
3. import/select the provided generalization hierarchies,
4. run risk + utility analysis, and
5. export a PDF report and an anonymized dataset.

⚠️ Do not use real or non de-identified patient data (no raw DICOM headers). This configuration assumes that DICOM images have already been de-identified.

---

1) Prerequisites (what you need installed)

Install ARX Data Anonymization Tool (GUI). Official pages:

https://arx.deidentifier.org/anonymization-tool/
https://arx.deidentifier.org/downloads/

That’s it — you do not need to build code from this repo just to use the hierarchies/templates.

---

2) What’s in this repository

Typical folders:

• hierarchies/
  Pre-built hierarchy files (CSV) for EUCAIM-relevant quasi-identifiers (QIs) — e.g., dates, nationality, manufacturer, alergies etc.

• examples/
  Small synthetic example datasets (CSV/XLSX) and example outputs (report/anonymized file).

• docs/
  User guidance: how to import data, assign attribute roles, load hierarchies, run risk/utility analysis, and interpret results.

---

3) Before ARX: prepare your input table

A. DICOM metadata

ARX works on tabular data, not on .dcm files directly.
First, extract the relevant DICOM tags into a .csv/.xlsx (one study row per record), then analyze that table in ARX.

B. Clinical / cohort data

Any cohort-style table works:

• rows = subjects/studies/records
• columns = demographics, lab values, conditions, therapies, imaging metadata fields, etc.

C. One cohort = one ARX project

Treat imaging metadata and clinical information as separate projects.

---

4) Run the Wizard workflow in ARX (step-by-step)

Step 1 — Create a new ARX project and import your CSV/XLSX

1. Open ARX
2. Create a new project
3. Import your dataset (.csv or .xlsx)

Step 2 — Assign attribute roles (this matters most)

For each column, decide whether it is:

• Identifier: direct identifiers (typically removed / not used for release)
• Quasi-Identifier (QI): enables re-identification when combined (needs generalization/suppression)
• Sensitive attribute: values you want to protect (e.g., diagnoses)
• Insensitive attribute: safe to keep as-is

This classification is central to defining the solution space ARX will explore.

Step 3 — Load/select hierarchies for QIs

For each QI column:

1. Import the matching hierarchy CSV from this repository, inside the /hierarchies folder
2. Confirm hierarchy levels are correct (from specific → more general)

Example hierarchy idea (date masking like YYYYMMDD → YYYYMM* → YYYY****)

Step 4 — Choose a privacy / risk model

Select an anonymization model appropriate to your context, commonly:

• k-anonymity (and extensions)
• optionally define limits for generalization/suppression
• optionally allow a small percentage of outliers to be ignored (if appropriate)

Step 5 — Run analysis and explore results

Run the analysis. Then use ARX’s Explore Results view to:

• compare candidate transformations
• filter by compliance (color-coded)
• understand which attributes drive risk and which drive utility loss

Step 6 — Utility analysis (don’t skip this)

After selecting a candidate solution, evaluate usefulness:

• generic information-loss metrics
• (optional) classification models to compare performance on original vs anonymized data

Step 7 — Export

Export:

• a PDF report describing the selected generalization levels and results, and
• the anonymized dataset (.csv/.xlsx)

---

5) Using (and editing) hierarchy files

Hierarchy files are typically CSV tables with multiple generalization levels per value.

Example concept (nationality, 3 levels):

• Level 0: specific value
• Level 1: broader grouping
• Level 2: coarsest grouping

Tips:

• Keep hierarchies consistent with your column’s value format
• Version-control hierarchy changes so analyses remain reproducible

---

6) Common problems & fixes

• “Invalid input dataset format”
  Check CSV delimiter/encoding, header row, and consistent column types.

• “Too strict anonymity requirements” / no feasible solutions
  Try one or more of:

  • relax k / privacy parameters (if allowed)
  • allow more suppression
  • use coarser hierarchies
  • reduce the number of QIs (only if justified)

• Utility loss is too high
  Consider alternative hierarchies, attribute weights, or different constraints; compare utility metrics and/or classification performance.

---