junocompanion/juno-symptom-taxonomy

Juno Symptom Taxonomy

A plain-language catalogue of 613 patient-reportable symptoms and experiences in 19 body-system groups. Each entry has a canonical name, a one-sentence everyday description and, where available, the words people actually use for it ("wiped out", "pins and needles", "racing heart").

The taxonomy was built by Marshall Gould for Juno, a conversational companion for people living with chronic illness. It is described in the paper below.

Gould M. Representing patient reported symptoms in conversational digital health: exploratory catalogue development and candidate retrieval study. Submitted to medRxiv (manuscript MEDRXIV/2026/364650, awaiting screening; DOI pending). PDF: https://junocompanion.com/chronic-illness-support/tracker/symptoms#research

Not for diagnosis. This is a vocabulary, not a symptom checker. See Limitations.

Try it: Juno Symptom Finder matches what you type to the closest entries.

Quick start

from datasets import load_dataset

taxonomy = load_dataset("junocompanion/juno-symptom-taxonomy", "taxonomy", split="train")
vocab = load_dataset("junocompanion/juno-symptom-taxonomy", "app_vocabulary", split="train")

print(taxonomy[0])

The same tables are available as CSV in data/ (list fields are joined with "; ").

What is in it

Config taxonomy (default, 613 rows)

The research catalogue that the paper describes, with restricted fields removed.

Field	Type	Description
id	string	Stable Juno identifier, for example current-brain-fog. Identifiers never change. The prefix is historical; current- entries are the names the Juno app uses today.
canonical_symptom	string	The display name. Specialist terms carry a plain gloss, for example Allodynia (pain from touch).
body_system	string	One of 19 browsing groups, for example Heart & circulation or Thinking & communication. Groups are for finding things, not anatomical claims.
plain_description	string	A one-sentence everyday description written by Juno.
consumer_phrases	list of strings	Everyday wordings and alternative names, written by Juno (see Provenance). 166 entries have at least one, 303 in total.
entry_type	string	Juno's recording class (see below).
reference_urls	list of strings	Public pages the description was checked against, mostly NHS, DermNet, Cleveland Clinic, NICE and patient charities. These are citations for the check. Text was not copied from them.

There is no red_flag field because the source catalogue does not have one. Use your own clinically governed safety rules.

Entry types:

entry_type	n	Meaning
symptom	323	Something a person feels or notices
finding	103	An observable change, for example a rash or swelling
clinical-topic	72	The name of a condition or clinical concept, for example Sleep apnoea. Not a symptom: keep these apart when you need symptoms only
episode	46	A bounded occurrence, for example a panic attack
experience	43	Something a person may want to record without implying illness, for example loneliness or déjà vu
state	10	A sustained state, for example autistic burnout
treatment-experience	9	Something felt in connection with a treatment
event, symptom-pattern, seizure-feature, treatment-event, care-routing-topic	7	Small specialised classes. care-routing-topic (1 entry) is a warning-sign topic that needs a separate care policy

Body-system groups and sizes:

body_system	n	body_system	n
Mood & emotional wellbeing	63	Temperature & whole-body changes	32
Muscles, joints & pain	59	Bladder & urinary health	26
Skin, hair & nails	57	Breathing & chest	25
Digestion & bowel	49	Heart & circulation	24
Headaches & nerve symptoms	47	Eyes & vision	23
Ears, nose & mouth	43	Sleep	23
Pelvic, reproductive & sexual health	40	Treatment effects	19
Movement & balance	35	Sensory experiences	16
Energy & recovery	13	Thinking & communication	16
Breast changes	3

Example row:

{
  "id": "current-palpitations",
  "canonical_symptom": "Palpitations",
  "body_system": "Heart & circulation",
  "plain_description": "An unusually noticeable heartbeat, such as pounding, fluttering or a feeling of skipped beats.",
  "consumer_phrases": ["heart palpitations", "heart racing", "palpitation", "pounding heart", "racing heart"],
  "entry_type": "symptom",
  "reference_urls": ["https://www.nhs.uk/conditions/heart-palpitations/"]
}

Config app_vocabulary (201 rows)

The compact vocabulary that the Juno app uses in production to store symptoms. Free text is folded onto one of these 201 lowercase names.

Field	Type	Description
id	string	app- plus the name as a slug
canonical_symptom	string	The stored name, for example shortness of breath
app_group	string	One of the app's 15 groups, for example Breathing, nose & throat
consumer_phrases	list of strings	Juno's phrase table for this name, for example breathless, short of breath
taxonomy_id	string	The matching entry in the taxonomy config. All 201 names are linked

Use taxonomy when you need breadth and descriptions. Use app_vocabulary when you want a small, stable label set for normalising free text.

Intended uses

• Symptom pickers and search in patient-facing apps, especially for chronic and multi-system conditions.
• Normalising free-text symptom mentions to a consumer-friendly concept list, then mapping those concepts to a clinical terminology under your own licence.
• Research on consumer health vocabulary and patient language, for example as a target label set for entity linking.
• Teaching and prototyping, as in the Juno Symptom Finder Space.

Limitations (not for diagnosis)

• Not a diagnostic or triage tool. Entries describe what people report. They say nothing about cause, severity or urgency. Do not use this dataset to suggest diagnoses, rule conditions out or decide whether someone needs care.
• No clinical validation. The descriptions and groupings are editorial design decisions. They were AI-assisted and source-checked, but no independent clinician panel adjudicated them. The paper reports development results only.
• No terminology codes. SNOMED CT, ICD and MedDRA content is deliberately excluded. Matching an entry is not a clinical code.
• Incomplete everyday language. Only 166 of 613 entries have consumer phrases. The phrases were written by Juno and have not been tested with patients. A missing match does not mean a symptom is unimportant or unrecognised.
• English only. Wording is mostly UK English in taxonomy and US English in app_vocabulary.
• Scope. The catalogue was designed for adults managing chronic conditions. Coverage of paediatric, obstetric and acute-care language is limited.
• Clinical topics. clinical-topic entries name conditions. Showing them as "matches" to a symptom description can read as a diagnosis, so filter them out of symptom-only interfaces.
• Snapshot. This is the catalogue export of 14 September 2026 (614 rows, 20 browsing groups) minus one interface-only row. The paper's frozen baseline has 612 active entries and its revised proposal has 615. The three entries added in that revision (Pain, Discomfort, Feeling empty) are not in this release.

How it was built

The paper describes the method in full. In summary:

1. Juno's catalogue grew from the symptom names used in the app. The version reviewed for the paper had 612 active entries in 20 browsing groups, linked to 80 condition packs.
2. An AI-assisted editorial and rule-based review gave each entry a plain-language definition, a recording class and a distinction from neighbouring entries. Definitions were checked against public patient-information sources, which are listed in reference_urls.
3. The paper then evaluated candidate retrieval against two public research corpora, COMETA and PriMock57. Those corpora were used only to evaluate retrieval. No text from them is in this dataset, and this snapshot was exported before the corpus-informed revisions.
4. For this release, the catalogue export was joined with the production app's vocabulary and phrase tables on the shared current- identifiers. Restricted fields were then removed (see below).

Provenance and licence

Licence: CC BY 4.0. Please credit "Juno symptom taxonomy, Marshall Gould / Juno" and cite the paper.

Sources (all Juno-authored):

• the research catalogue export behind the paper (names, groups, definitions, recording classes and aliases)
• the Juno app's production vocabulary and phrase tables (vocabulary.json, synonyms.ts and the picker aliases in symptom-catalog.ts)

Removed before release:

What	Why
SNOMED CT concept IDs, fully specified names and all mapping metadata (relationship, status, edition, version)	SNOMED CT content is licensed by SNOMED International and cannot be redistributed
5 reference links to terminology-server records of SNOMED CT concepts	They point to licensed terminology content
Symptom-to-condition links and editorial notes on them	Most associations are marked as needing evidence, and they would invite diagnostic use
Icon paths, review statuses, merge history	Internal app data, not taxonomy content
The "Other / your own symptom" row	An interface action, not a symptom
9 vague-word routing rules (for example pain → body aches)	App routing decisions, not synonyms
8 site-specific aliases (for example numb feet) that the app folds onto a broader name	They name a body site, not the broader concept
5 phrases that are the name of a different entry (for example period pain)	To avoid pointing one name at two entries

Not included at all: any text from COMETA, PriMock57, CADEC or other corpora; ICD or MedDRA text; model weights; and any user data. No user-entered text, conversation or account-level data was used to make this release.

Many entry names are ordinary English medical words, such as "Ataxia" or "Headache", that also appear in clinical terminologies. They are Juno's labels, released without any terminology identifiers or descriptions.

Not in this release

The paper's data statement also mentions mapping files, editorial annotations and analysis code. They are not released here:

• Terminology mapping files consist of SNOMED CT identifiers and descriptions, which cannot be redistributed.
• Editorial annotations quote or label records from COMETA and PriMock57, which are not ours to redistribute.
• Analysis code for the retrieval experiments is being recovered and may follow in the GitHub repository.

Citation

Paper (preprint, DOI pending):

@misc{gould2026representing,
  author       = {Gould, Marshall},
  title        = {Representing patient reported symptoms in conversational digital health: exploratory catalogue development and candidate retrieval study},
  year         = {2026},
  howpublished = {medRxiv preprint, submitted (manuscript MEDRXIV/2026/364650), awaiting screening},
  note         = {DOI pending. PDF: https://junocompanion.com/chronic-illness-support/tracker/symptoms\#research},
  url          = {https://junocompanion.com/chronic-illness-support/tracker/symptoms#research}
}

Dataset:

@misc{juno_symptom_taxonomy_2026,
  author       = {Gould, Marshall},
  title        = {Juno Symptom Taxonomy},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/junocompanion/juno-symptom-taxonomy}},
  note         = {Version 1.0, CC BY 4.0}
}

Links

• Juno: https://junocompanion.com
• Research page and paper PDF: https://junocompanion.com/chronic-illness-support/tracker/symptoms#research
• Code and release mirror: https://github.com/MarshallBear1/juno-symptom-taxonomy
• Author: Marshall Gould, Juno (SharedGenes, Inc.), ORCID 0009-0006-8769-1853. Contact: team@juno-chat.com