README.md

---
license: cc-by-nc-sa-4.0
language:
- ru
multilinguality:
- monolingual
annotations_creators:
- expert-generated
source_datasets:
- original
task_categories:
- token-classification
task_ids:
- named-entity-recognition
pretty_name: RuClinNER
size_categories:
- 1K<n<10K
tags:
- clinical
- medical
- biomedical
- named-entity-recognition
- nested-ner
- relation-extraction
- entity-linking
- icd-10
- russian
configs:
- config_name: medterm
  data_files:
  - split: train
    path: data/medterm/train.parquet
  - split: test
    path: data/medterm/test.parquet
- config_name: drugrel
  data_files:
  - split: train
    path: data/drugrel/train.parquet
  - split: test
    path: data/drugrel/test.parquet
- config_name: covid_rmj
  data_files:
  - split: train
    path: data/covid_rmj/train.parquet
  - split: test
    path: data/covid_rmj/test.parquet
---

# RuClinNER

RuClinNER is a collection of three complementary Russian **clinical** named-entity-recognition
corpora. They span distinct kinds of clinical text written in healthcare settings —
patient-reported complaints, physician prescriptions, and peer-reviewed clinical research
articles — and together contain **55,185 entity spans across 22 entity types** and
**18,112 relations across 8 relation types** over **6,797 records**.

The collection accompanies the paper *RuClinNER: Russian Clinical Multi-Label NER Annotated
Corpora*. It is intended for research on Russian clinical and biomedical NLP, where annotated
clinical text remains scarce: existing Russian resources largely target consumer drug reviews
or scientific abstracts rather than text produced in clinical practice.

## The three corpora

| Config | Text type | Records | Entities | Types | Relations | Median words | Notable |
|---|---|--:|--:|--:|--:|--:|---|
| `medterm`   | patient complaints   | 4,474 | 12,582 | 3  | —      | 9     | ICD-10 normalization subset |
| `drugrel`   | drug prescriptions   | 2,200 | 35,635 | 13 | 18,112 | 28    | nested entities + relations; patient metadata |
| `covid_rmj` | clinical journal articles | 123 | 6,968 | 6 | —  | 2,386 | long documents; dual annotation layers |
| **Total**   |                      | **6,797** | **55,185** | **22** | **18,112** | | |

- **MedTerm** — short, free-text symptom narratives from a symptom-checking service.
  Three entity types (`SYMPTOM`, `RISK`, `DISEASE`). A curated subset of entities is linked
  to **ICD-10** canonical concepts (Russian and English names), supporting clinical concept
  normalization / entity linking.
- **DrugRel** — Russian clinical prescription records. Thirteen entity types covering drug
  names, dosage components, and modifiers, with **hierarchically nested** entities: a `REGIME`
  span is a *container* whose tokens also carry nested labels (`DOSE`, `DOSIR`, `FREQ`, `DUR`,
  `TIME`, `USE`, …). Eight relation types link attributes to their drug. Records are linked to
  de-identified patient metadata (age, gender, ICD-10 diagnosis codes).
- **COVID-RMJ** — full-length COVID-19 articles from the Russian Medical Journal, annotated by
  medical experts over six entity types. Two parallel annotation layers are provided: the full
  article text and a structured abstract, each with character spans and pre-computed BIO tags.

## What the data looks like

### MedTerm — patient complaint

A single short complaint often mixes all three entity types:

> зевота, нехватка кислорода, астма, аллергия

| span | label |
|---|---|
| зевота | `SYMPTOM` |
| нехватка кислорода | `SYMPTOM` |
| астма | `DISEASE` |
| аллергия | `RISK` |

**ICD-10 normalization.** A curated MedTerm subset links spans to canonical concepts.
Patients write the same concept many ways — *Fever* (`R50.9`) alone appears as, among 50+ forms:

> температура · Температура 38 · t 37 · 37,4 · до 39 · лихорадка · температкра · Темперара

(the last two are genuine patient typos) — all normalized to **Fever, unspecified — R50.9**.

### DrugRel — prescription with nested entities and relations

A `REGIME` is a *container* span: the tokens inside it also carry nested labels, and each
attribute is linked back to its drug by a typed relation.

> Левофлоксацин по 500 мг 1 раз в день 5 дней

- `DRUG` — **Левофлоксацин**
- `REGIME` — **по 500 мг 1 раз в день 5 дней**  *(container — its tokens carry a second label)*
  - `DOSE` — по 500 мг
  - `FREQ` — 1 раз в день
  - `DUR` — 5 дней
- Relations: `DOSE → DRUG`, `REGIME → DRUG`

### COVID-RMJ — clinical journal article

Dense, expert-annotated entities over long documents:

> после положительного результата **ПЦР** на **SARS-CoV-2** различные симптомы
> (**усталость** — 39,0 %, **головная боль** — 23,2 %, **одышка** — 23,4 % и др.)

| span | label |
|---|---|
| ПЦР | `PROCEDURE` |
| SARS-CoV-2 | `DISEASE` |
| усталость · головная боль · одышка | `SYMPTOM` |

## Supported tasks

The corpora support several clinical-NLP tasks:

- **Named entity recognition** — all three corpora (character-offset spans).
- **Nested / multi-label token classification** — DrugRel `REGIME` containers (≈22,900
  overlapping span pairs); a single token may carry more than one label.
- **Relation extraction** — DrugRel drug–attribute relations (8 types).
- **Clinical concept normalization / entity linking** — MedTerm entities linked to ICD-10.
- **Long-document NER** — COVID-RMJ articles (median ≈2,400 words).
- **Cross-corpus and domain-transfer studies**, and **evaluation of LLM prompting** for
  clinical information extraction across complementary text types.

## Usage

```python
from datasets import load_dataset

medterm = load_dataset("medlinx/RuClinNER", "medterm")
drugrel = load_dataset("medlinx/RuClinNER", "drugrel")
covid   = load_dataset("medlinx/RuClinNER", "covid_rmj")

ex = medterm["test"][0]
print(ex["text"])
for e in ex["entities"]:
    print(e["label"], ":", e["text"], e.get("coding"))
```

## Data fields

**`medterm`**
- `id` (string), `text` (string)
- `entities`: list of `{start, end, label, text, coding}` where `coding` is
  `{icd10, name_ru, name_en}` for the normalized subset, else `null`.

**`drugrel`**
- `id`, `record_id` (string), `text` (string)
- `entities`: list of `{id, start, end, label, text}`
- `relations`: list of `{type, from_id, to_id, from_text, to_text}`
- `patient_age` (int, nullable), `patient_gender` (string, nullable)
- `diagnosis` (string), `diagnosis_icd_codes` (list of string)

**`covid_rmj`**
- `id`, `text` (string)
- `text_entities`: list of `{start, end, label, text}` over `text`
- `text_tokens`, `text_ner_tags`: pre-tokenized text with BIO tags
- `annotation` (string, structured abstract) with `annotation_entities`,
  `annotation_tokens`, `annotation_ner_tags`

## Splits

| Config | train | test |
|---|--:|--:|
| `medterm`   | 3,049 | 1,425 |
| `drugrel`   | 2,100 | 100 |
| `covid_rmj` | 98 | 25 |

For MedTerm, the ICD-10 normalization subset (508 records, 1,294 linked entities) falls within
the `test` split. For COVID-RMJ, a train/test split is provided here; because the corpus is
small (123 documents), the accompanying paper instead evaluates encoders with 5-fold
cross-validation.

## Entity and relation types

**MedTerm** — `SYMPTOM` (symptoms/signs), `RISK` (risk factors), `DISEASE` (named conditions).

**DrugRel** — `DRUG`, `REGIME` (full dosage regimen; container), `DOSE` (numeric dose),
`DOSIR` (per-intake instruction), `FREQ`, `DUR`, `TIME`, `USE` (route), `FORM`, `COND`
(condition for taking), `INN` (active substance), `NUM`, `NEG`. Relations:
`REGIME-DRUG`, `FORM-DRUG`, `DOSE-DRUG`, `USE-DRUG`, `COND-DRUG`, `INN-DRUG`, `NEG-DRUG`, `ALT`.

**COVID-RMJ** — `DISEASE`, `PROCEDURE`, `SYMPTOM`, `DRUG`, `ANATOMY`, `LAB_VALUE`.

## Dataset creation

Each corpus was annotated by domain experts (medical informaticians, clinical pharmacologists,
and medical specialists) following written guidelines. Records were independently annotated and
disagreements adjudicated. Inter-annotator agreement (Cohen's κ): MedTerm 0.87 (boundaries) /
0.91 (labels); DrugRel 0.89 (entities) / 0.81 (relations). COVID-RMJ was reviewed by at least
two experts with adjudication and passes automated boundary/overlap/BIO consistency checks.

## Considerations for using the data

- **Class imbalance** is substantial and corpus-specific (e.g. `SYMPTOM` ≈85% of MedTerm
  spans; `DISEASE` ≈61% of COVID-RMJ spans). Report macro- as well as micro-averaged metrics.
- **MedTerm** contains genuine negatives: ≈18% of records have no annotated entity.
- **DrugRel** includes 67 relations with `type=null` (annotation artifacts); exclude these for
  relation-extraction evaluation. The `RISK`/`DISEASE` distinction can be semantically
  ambiguous (e.g. obesity).
- **DrugRel demographics** skew toward female patients (≈74%), with many pediatric and
  obstetric records; consider this when transferring models to other populations.
- **COVID-RMJ** entity density varies ~100× across documents.

## Personal and sensitive information

The data is de-identified: MedTerm and DrugRel records contain no patient names, dates of
birth, or contact information; COVID-RMJ articles are published, peer-reviewed texts with no
patient-level data. DrugRel patient metadata is limited to coarse age, gender, and ICD-10
diagnosis codes.

## Licensing

Released under **Creative Commons Attribution-NonCommercial-ShareAlike 4.0 (CC BY-NC-SA 4.0)**
for research use.

## Citation

```bibtex
@misc{ruclinner2026,
  title        = {RuClinNER: Russian Clinical Multi-Label NER Annotated Corpora},
  author       = {Gudkov, Vadim and Varennikova, Anastasia and Miroshnichenko, Polina and
                  Mitrofanova, Olga and Gousyatskaya, Polina and Boitsova, Daria},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/medlinx/RuClinNER}}
}
```

## Contact

Maintained by [Medlinx](https://huggingface.co/medlinx). Questions: `gudkov.v@medlinx.online`.
</content>