SPARCLING-DE: Synthetic German Clinical Discharge Letters with PHI Labels by Construction

SPARCLING-DE is a benchmark and training dataset of 1,019 synthetic German hospital discharge letters (Arztbriefe) containing 837,713 tokens and 34,712 character-exact Protected Health Information (PHI) span annotations across 25 PHI classes and 3 non-PII negative trap classes.

It accompanies the paper:

Generating Diverse and Verified Synthetic Clinical Notes with Labels by Construction
Matthias Lehmann, Madeleine Werthebach
GenAI for Health Workshop @ NeurIPS 2026
Code: https://github.com/Matt00n/phi_discharge_letter_generator

---

Dataset Summary

Evaluating and training clinical de-identification models requires benchmarks with:

1. Exact and complete character-level PHI span annotations
2. Surface plausibility sufficient to embed PHI in realistic clinical context
3. Pervasive yet controlled diversity across document styles, clinical specialties, demographic distributions, and adversarial edge cases.

SPARCLING (Synthetic, Provenance-Annotated, Red-teamed, CLINical data Generation engine) generates discharge letters through a four-stage pipeline:

1. Structured Attribute Sampling: Patient demographics, clinical cases (3,046 unique ICD-10-GM codes), document styles, and adversarial configurations are sampled hierarchically and deterministically from curated distributions ($z_i \sim P(z)$).
2. Controlled Template Authoring: A masked specification ($\tilde{z}_i$) is passed to an LLM (gemini-3.7-flash), which authors a clinical narrative using typed, grammatically inflected placeholders ([[ref form=... case=...]]) without ever seeing raw PHI values.
3. Deterministic Rendering: Placeholders are deterministically substituted with morphologically inflected values while tracking exact character offsets $(x_i, Y_i) = g_{\text{fill}}(t_i, z_i)$, yielding character-exact gold spans by construction.
4. Dual Quality Control: Every generated document is audited by (a) a deterministic known-PHI leak and negative-trap scanner and (b) an adversarial red-teaming LLM on the masked text to catch any hallucinated residual PHI. Any flagged document is excluded.

---

Configurations & Files

This repository provides two Hugging Face datasets configurations as well as the full provenance artifacts:

File / Config	Records	Description
dataset.jsonl (default config)	1,019	Full verified SPARCLING-DE benchmark and training corpus
subsample63.jsonl (subsample63 config)	63	Size-matched representative subsample (matching GraSCCo's $N=63$ for corpus-size-sensitive lexical Comparisons)
review_queue.jsonl	114	Documents excluded by Stage-4 Dual Quality Control (for QC analysis / Ablation 1)
failures.jsonl	17	Stage 1–3 generation/rendering drops
provenance.json	—	Full yield chain, exclusion counts by mechanism, deterministic vs. red-team overlap
diversity.json	—	6-dimension diversity audit report
subsample63_meta.json	—	Total-variation distance (TVD) and Kolmogorov–Smirnov (KS) representativeness checks for subsample63

Quick Load

from datasets import load_dataset

# Load full SPARCLING-DE corpus (1,019 documents)
ds = load_dataset("Matt00n/SPARCLING-DE", name="default", split="train")

# Load the N=63 representative subsample
ds_63 = load_dataset("Matt00n/SPARCLING-DE", name="subsample63", split="test")

To use directly with the SPARCLING evaluation & benchmark repo, download the files into experiments/datasets/sparcling-de/:

huggingface-cli download Matt00n/SPARCLING-DE \
  --repo-type dataset \
  --local-dir experiments/datasets/sparcling-de

---

Data Schema

Each record in dataset.jsonl and subsample63.jsonl has the following structure:

{
  "doc_id": "doc_00000",
  "text": "Universitätsklinikum Berg\nKlinik für Gastroenterologie...",
  "spans": [
    {
      "start": 0,
      "end": 25,
      "label": "INSTITUTION",
      "tier": "QUASI",
      "entity_id": "institution",
      "form": "default",
      "case": null
    }
  ],
  "negatives": [
    {
      "start": 2201,
      "end": 2215,
      "label": "EPONYM",
      "tier": "NONPII",
      "entity_id": "eponym_2",
      "form": null,
      "case": null
    }
  ],
  "metadata": {
    "doc_id": "doc_00000",
    "seed": 1000000,
    "style": "Standard-Klinik",
    "difficulty": "hard",
    "difficulty_tags": ["needle", "quasi"],
    "sections": ["Anamnese", "Diagnosen", "Aufnahmebefund", "..."],
    "fachabteilung": "Innere Medizin – Gastroenterologie",
    "patient_surname": "Çam",
    "patient_origin": "turkish",
    "patient_age": 62,
    "marital_status": "verheiratet",
    "pii_density": 0.642,
    "code_style": "none",
    "opening_style": "bericht",
    "framing": {
      "briefkopf": true,
      "anrede": true,
      "patientenstammdaten": true,
      "schlussformel": true
    },
    "has_medications": true,
    "n_spans": 38,
    "n_negatives": 4,
    "quasi": ["beruf", "alter"],
    "splits": {
      "by_domain": "train",
      "by_difficulty": "hard",
      "by_surface": "train"
    }
  },
  "qc": {
    "flagged": false,
    "reasons": []
  }
}

---

Label Ontology (25 PHI Labels + 3 Non-PII Negative Traps)

Tier	Family	Labels
Direct Identifiers (17)	Names	PATIENT_NAME, RELATIVE_NAME, STAFF_NAME, OTHER_PERSON_NAME
	Dates	DATE_BIRTH, DATE_ADMISSION, DATE_DISCHARGE, DATE_OTHER
	Location	ADDRESS
	Contact	PHONE, FAX, EMAIL, URL
	Identifiers	ID_INSURANCE (KVNR), ID_CASE (Fallnr.), ID_PATIENT, ID_OTHER (LANR/BSNR/IK)
Quasi-Identifiers (8)	Location	ZIP, CITY
	Organizations	INSTITUTION, INSURER
	Demographics	AGE, PROFESSION, EMPLOYER, NATIONALITY
Non-PII Negatives (3)	Decoy Traps	EPONYM (e.g., Morbus Crohn), LOOKALIKE (e.g., Swan-Ganz-Katheter), NONPII_ID (e.g., PZN)

Important: negatives are non-PHI clinical terms and codes that resemble personal identifiers. They must not be redacted and are used to measure clinical over-redaction (false positives).

---

Ethical Considerations

All documents and patient profiles in SPARCLING-DE are 100% synthetically generated from scratch using curated distributions and public medical catalogs (ICD-10-GM, OPS). No real patient data or clinical records were used at any stage. Because generation prioritizes benchmark diversity and stress-testing edge cases over deep clinical validity, the texts should not be used as a source of clinical medical knowledge.

---

Citation

@inproceedings{lehmann2026sparcling,
  title     = {Generating Diverse and Verified Synthetic Clinical Notes with Labels by Construction},
  author    = {Lehmann, Matthias and Werthebach, Madeleine},
  booktitle = {GenAI for Health Workshop at NeurIPS 2026},
  year      = {2026}
}