README.md

---
license: cc-by-4.0
language:
  - en
task_categories:
  - token-classification
task_ids:
  - named-entity-recognition
tags:
  - medical
  - clinical-nlp
  - ner
  - information-extraction
  - synthetic
  - icd10
  - clinical-ips
pretty_name: SynthIPS
size_categories:
  - n<1K
configs:
  - config_name: documents
    data_files: data/documents/*.parquet
    default: true
  - config_name: chunks
    data_files: data/chunks/*.parquet
  - config_name: patient_summary
    data_files: data/patient_summary/*.parquet
---

# SynthIPS — Synthetic IPS NER Validation Dataset

A validation dataset of **10 synthetic patients** across **29 clinical encounters**, annotated for the three core entity types of the **International Patient Summary (IPS)** standard: **Condition**, **Medication**, and **Lab**.

The [International Patient Summary (IPS)](https://www.hl7.org/fhir/uv/ips/) is an HL7 FHIR standard for a minimal, specialty-agnostic health record designed for cross-border and cross-system care continuity. SynthIPS is built around its three primary clinical sections — problems, medications, and lab results — and provides realistic clinical documents with ground-truth annotations aligned to IPS entity types, ICD-10 condition codes, ATC medication codes, and LOINC lab identifiers.

Documents include structured blocks (problem lists, medication tables, lab grids) alongside free-text narrative sections and deliberate distractor spans (negated / ruled-out entities), reflecting the mixed layout of real clinical encounter documents.

| Stat | Value |
|------|-------|
| Patients | 10 |
| Encounters | 29 |
| GT spans | 171 (86 Condition · 85 Medication) |
| Distractor spans | 14 |
| Structured lab analytes | 446 (35 unique, 100% LOINC coverage) |
| Text segments | 815 (111 contain GT entities) |
| Encounter types | initial\_consult · follow\_up · specialist\_letter · acute\_visit · lab\_review |
| Patient profiles | minimal · typical · complex (controls entity density) |

---

## Configs

All configs share `doc_id` (e.g. `"patient_01/encounter_01"`) and `patient_id` as join keys.

| Config | Rows | Primary use |
|--------|------|-------------|
| `documents` *(default)* | 29 | Full archive — document text, PDF, all GT annotations |
| `chunks` | 815 | NER evaluation — one row per document segment with inline entity annotations |
| `patient_summary` | 10 | Patient-level overview — aggregated GT spans, labs, encounter list |

### How configs relate

```
patient_summary ── patient_id ──► documents ◄── doc_id ──► chunks
                                  (source of truth)
```

**Lab data** lives in `documents.structured_entities` (filter to `block == "lab_grid"`).

---

## Quick start

```python
from datasets import load_dataset
import json

# ── Chunk-level NER (recommended starting point) ─────────────────────────
chunks = load_dataset("betterInnovation/SynthIPS", "chunks", split="test")

for chunk in chunks:
    entities = json.loads(chunk["gt_entities"])
    if not entities:
        continue
    for e in entities:
        span = chunk["markdown_text"][e["start"]:e["end"]]
        print(chunk["segment_type"], f"p{chunk['page_number']}",
              e["label"], repr(span), "|", e["kg_id"])

# ── Full-document NER (join chunks → documents for full text) ─────────────
docs = load_dataset("betterInnovation/SynthIPS", "documents", split="test")
doc = docs[0]

gt_spans = json.loads(doc["gt_spans"])
for e in gt_spans:
    span = doc["markdown_text"][e["start"]:e["end"]]
    print(e["surface"], "→", e["label"], "|", e["canonical"])

# ── Lab data ──────────────────────────────────────────────────────────────
labs = [e for e in json.loads(doc["structured_entities"])
        if e["block"] == "lab_grid"]
for lab in labs:
    print(lab["analyte"], lab["value"], lab["unit"], lab["flag"], lab["loinc"])

# ── PDF bytes (rendered natively in HF Data Viewer) ───────────────────────
# doc["pdf_bytes"] → {"bytes": b"...", "path": None}  (~70 KB, 3 pages each)

# ── Patient summary ───────────────────────────────────────────────────────
patients = load_dataset("betterInnovation/SynthIPS", "patient_summary", split="test")
for p in patients:
    print(p["patient_id"], p["profile"],
          p["n_encounters"], "encounters")
    print("  encounter types:", list(p["encounter_kinds"]))  # native Arrow list
```

---

## Column reference

### `documents` config — 29 rows

One row per encounter. The default config. PDF renders natively in the HF Data Viewer.

| Column | Type | Description |
|--------|------|-------------|
| `doc_id` | string | `"patient_01/encounter_01"` — join key |
| `patient_id` | string | `"patient_01"` — join key |
| `pdf_bytes` | Pdf | Raw PDF (~70 KB, 3 pages). Renders as page viewer in HF Data Viewer. |
| `markdown_text` | string | Full document text as markdown — header, problem list, medication table, lab grid, narrative, signatures. `start`/`end` offsets in `gt_spans` point here. |
| `patient_name` | string | |
| `patient_profile` | string | `minimal / typical / complex` |
| `patient_sex` | string | `"M"` / `"F"` |
| `patient_dob` | string | ISO date |
| `anchor_icd10` | string | Primary ICD-10 codes |
| `clinical_hook` | string | One-line clinical summary |
| `encounter_id` | string | `"encounter_01"` |
| `encounter_kind` | string | `initial_consult / follow_up / specialist_letter / acute_visit / lab_review` |
| `encounter_title` | string | Full encounter title |
| `encounter_date` | string | ISO date |
| `provider_name` | string | |
| `provider_role` | string | |
| `provider_facility` | string | |
| `gt_spans` | string (JSON) | GT entity spans — `start`/`end` into `markdown_text` |
| `distractor_spans` | string (JSON) | Trap spans (negated / family history / hypothetical) — should NOT be extracted |
| `structured_entities` | string (JSON) | Structured block entities (lab grid, problem list, medications) with LOINC/ICD-10/ATC codes |

### `chunks` config — 815 rows

One row per document segment. `markdown_text` here is the segment's own slice of the full document text. Join on `doc_id → documents` for complete document context.

| Column | Type | Description |
|--------|------|-------------|
| `doc_id` | string | Join key → `documents` |
| `patient_id` | string | |
| `markdown_text` | string | This segment's text. `start`/`end` in `gt_entities` are offsets into this field. |
| `encounter_id` | string | |
| `encounter_kind` | string | |
| `patient_profile` | string | |
| `chunk_id` | string | Segment identifier within the encounter |
| `page_number` | int | Document page (1-indexed) |
| `chunk_order` | int | Reading order within the encounter |
| `segment_type` | string | `"text"` or `"table"` |
| `gt_entities` | string (JSON) | GT entities in this segment — `start`/`end` into `markdown_text` |
| `distractor_entities` | string (JSON) | Distractor spans in this segment |

### `patient_summary` config — 10 rows

One row per patient. Cross-encounter data pre-aggregated.

| Column | Type | Description |
|--------|------|-------------|
| `patient_id` | string | |
| `patient_name` | string | |
| `sex` | string | `"M"` / `"F"` |
| `dob` | string | ISO date |
| `profile` | string | `minimal / typical / complex` |
| `anchor_icd10` | string | Primary diagnoses |
| `clinical_hook` | string | One-line clinical summary |
| `n_encounters` | int | |
| `encounter_ids` | list[string] | Native Arrow list — no `json.loads` needed |
| `encounter_kinds` | list[string] | Native Arrow list |
| `all_gt_spans` | string (JSON) | All GT spans across all encounters — filter by `label` for Condition / Medication |
| `all_distractors` | string (JSON) | All distractor spans |
| `all_labs` | string (JSON) | All structured lab analytes across all encounters |

---

## Entity span schema

Applies to `gt_spans` / `distractor_spans` (documents) and `gt_entities` / `distractor_entities` (chunks).

| Field | Type | Description |
|-------|------|-------------|
| `surface` | str | Exact mention text as it appears in the document (may be abbreviation: `"HTN"`, `"T2DM"`) |
| `label` | str | IPS type: `Condition` or `Medication` |
| `start` | int | Start char offset into `markdown_text` (document-level in `documents`, segment-level in `chunks`) |
| `end` | int | End char offset into `markdown_text` |
| `kg_id` | str | ICD-10 code (Condition) or ATC code (Medication) |
| `canonical` | str | Normalised entity name — use for normalisation scoring |
| `section_kind` | str | Source section: `hpi` · `social_history` · `clinical_observation` · `assessment` |

**Offset guarantee:**
`markdown_text[start:end]` matches `surface` after whitespace normalisation.
Line-wrapping may insert `\n` between words — normalise with `re.sub(r'\s+', ' ', span)` before comparing.

**Distractors** share the same schema. They are contextually plausible entities that appear in the narrative but are *not* part of the patient's confirmed record (negated diagnoses, family history mentions, hypothetical treatments). A model that extracts them is producing false positives.

---

## Usage guide

### 1 · Chunk-level NER (primary track)

```python
import json
from datasets import load_dataset
from nervaluate import Evaluator

chunks = load_dataset("betterInnovation/SynthIPS", "chunks", split="test")

all_gold, all_preds = [], []
for chunk in chunks:
    gold = [
        {"label": e["label"], "start": e["start"], "end": e["end"]}
        for e in json.loads(chunk["gt_entities"])
    ]
    preds = your_model(chunk["markdown_text"])  # same format
    all_gold.append(gold)
    all_preds.append(preds)

evaluator = Evaluator(all_gold, all_preds, tags=["Condition", "Medication"])
results, _ = evaluator.evaluate()
```

### 2 · Full-document NER

```python
docs = load_dataset("betterInnovation/SynthIPS", "documents", split="test")

for doc in docs:
    gold = [
        {"label": e["label"], "start": e["start"], "end": e["end"]}
        for e in json.loads(doc["gt_spans"])
    ]
    preds = your_model(doc["markdown_text"])
    # ... accumulate and evaluate
```

### 3 · Distractor robustness (false-positive rate)

```python
from rapidfuzz import fuzz

def fires_on(pred_surface, distractor_surface, threshold=85):
    return fuzz.token_set_ratio(pred_surface.lower(), distractor_surface.lower()) >= threshold

for doc in docs:
    distractors = json.loads(doc["distractor_spans"])
    preds = your_model(doc["markdown_text"])
    fp = sum(
        1 for d in distractors
        if any(fires_on(p["text"], d["surface"]) for p in preds)
    )
    fp_rate = fp / len(distractors) if distractors else 0
```

### 4 · Lab analyte extraction

```python
for doc in docs:
    labs = [e for e in json.loads(doc["structured_entities"])
            if e["block"] == "lab_grid"]
    # Each lab: analyte, value, unit, flag, loinc
    # Score: precision/recall at analyte-name level against model output on markdown_text
```

### 5 · Entity normalisation scoring

Compare model output against `canonical` (normalised name) and `kg_id` (ICD-10 / ATC code).

---

## The 10 patients

| ID | Name | Sex | DoB | Profile | Primary conditions |
|----|------|-----|-----|---------|-------------------|
| patient\_01 | Donald Wilson | M | 1957-01-14 | typical | T2DM, hypertension, dyslipidaemia |
| patient\_02 | Kimberly Hernandez | F | 2000-01-31 | minimal | Asthma, allergic rhinitis |
| patient\_03 | Charles Lopez | M | 1943-07-09 | complex | Heart failure, arrhythmia, T2DM |
| patient\_04 | Jennifer Martin | F | 1978-11-23 | typical | MDD, GAD |
| patient\_05 | Patricia Moore | F | 1965-06-19 | typical | Knee OA, chronic low back pain, dyslipidaemia |
| patient\_06 | Karen Thomas | F | 1972-06-14 | minimal | Primary hypothyroidism |
| patient\_07 | William Johnson | M | 1947-07-17 | complex | COPD, hypertension, smoking |
| patient\_08 | Thomas Lee | M | 1955-11-14 | complex | Diabetic nephropathy CKD3, prior MI, angina |
| patient\_09 | Emily Hernandez | F | 1989-11-23 | typical | GERD, anxiety |
| patient\_10 | Karen Wilson | F | 1964-03-29 | typical | Osteoporosis, RA, hypothyroidism |

**Profile** controls entity density: `minimal` (1–3 problems/meds) · `typical` (3–6) · `complex` (6–12).

---

## Design notes

### IPS entity types

SynthIPS covers the three primary sections of the **IPS standard**:

| IPS section | Label in dataset | Identifier |
|-------------|-----------------|------------|
| Problems (conditions, diagnoses) | `Condition` | ICD-10 via `kg_id` |
| Medications | `Medication` | ATC code via `kg_id` |
| Results (lab analytes) | `Lab` | LOINC code via `loinc` field in `structured_entities` |

Entity names in `canonical` follow the IPS-preferred normalised form.
Abbreviations and synonyms (e.g. `"HTN"`, `"T2DM"`) are preserved in `surface` exactly as they appear in the document.

### Document structure

Each encounter document contains several distinct sections: institution header, patient demographics, structured problem list, medication table, structured lab grid, and free-text narrative sections (HPI, social history, clinical examination, assessment & plan). This mixed structure — combining tabular and narrative content — reflects the layout of real clinical encounter documents and is a deliberate challenge for NER models.

`markdown_text` preserves this structure as formatted markdown. All `start`/`end` offsets in entity spans point into `markdown_text`.

### Distractor design

Each encounter includes 0–1 deliberate trap spans: clinically plausible entities that appear in the narrative but are negated, belong to family history, are ruled out, or are hypothetical. Examples: *"no evidence of X"*, *"father had X"*, *"we will consider X if..."*. These measure whether a model extracts *only* confirmed, patient-present entities.

### Reproducibility

Every encounter document is deterministic given its `layout_seed`. Re-running the composer with the same source files produces byte-identical documents and GT annotations.

---

## Citation

```bibtex
@dataset{synthips2026,
  title   = {SynthIPS: Synthetic IPS NER Validation Dataset},
  author  = {Better Care},
  year    = {2026},
  url     = {https://huggingface.co/datasets/betterInnovation/SynthIPS},
  license = {CC BY 4.0}
}
```

---

## License

[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) — free to use, share, and adapt with attribution.