mkurman/clinical-case-icd10-diagnosis

Clinical History -> ICD-10 (Acute / Chronic) — CC BY-enriched

1798 de-identified clinical histories drawn from open-access case reports in PubMed Central, each paired with a single principal-diagnosis label: an ICD-10-CM code, its official descriptor, and an ACUTE/CHRONIC acuity status.

• —input: a de-identified clinical history (presentation only; the diagnosis is removed and no PHI is present).
• —output: {icd10_code, name, status} where status is ACUTE or CHRONIC (how the condition presented in that case).

Cases are distinguished by case_id prefix:

• —ED#### (320 cases) — emergency-medicine presentations (mostly acute).
• —CHR#### (360 cases) — chronic-disease case reports (mixed acute/chronic).
• —CC#### (1118 cases) — a CC BY-licensed expansion across ~60 disease areas, added to raise the commercially-reusable share.

Overall acuity: 1309 ACUTE / 489 CHRONIC.

Configs (by reuse license)

Three configurations, selected on the commercial-use axis of each source article's license:

from datasets import load_dataset

# only commercially-reusable cases (1530 rows)
ds = load_dataset("mkurman/clinical-case-icd10-diagnosis", "commercial", split="train")

# everything (default, 1798 rows)
ds = load_dataset("mkurman/clinical-case-icd10-diagnosis", split="train")
ds = load_dataset("mkurman/clinical-case-icd10-diagnosis", "non_commercial", split="train")

Licensing note. Each row carries its own license, license_url, and license_category. The commercial config contains only articles whose license permits commercial use; non_commercial is CC-NonCommercial. default additionally includes 19 cases whose source is under publisher copyright or an undetermined license — verify the per-row `license` field before redistributing those. Note that CC BY-ND / CC BY-NC-ND are no-derivatives licenses: they are grouped by commercial permission here, but assess whether your use constitutes a derivative.

Fields

Provenance & method

Cases were retrieved from PubMed Central / Europe PMC, full text isolated to the case-presentation section, then de-identified and labelled with an LLM. Every ICD-10 code was validated against the NLM Clinical Tables ICD-10-CM API and repaired to the nearest billable code where needed, with per-case anatomical/laterality checks against the case text. A second LLM pass audited each input for diagnosis leakage. The CC BY expansion was retrieved with an explicit LICENSE:"cc by" filter at Europe PMC, so its provenance license is authoritative.

Caveats

• —Codes are LLM-assigned and validated for billable validity, not audited by a certified coder; a small number are best-fit approximations where no exact-anatomy billable code exists.
• —Derived from published case reports (often unusual presentations) — not representative of routine epidemiology.
• —De-identified from already-public text; no protected health information is introduced.