ghanaopenai/ghana-health-symptoms

Ghana Health Symptoms -- Twi patient descriptions with English triage tags

98,656 first-person symptom descriptions in Twi, each labelled with one fine-grained English phrase naming the presenting problem, grounded in 36,565 passages of real Twi health speech.

Every row is what a patient might say when a nurse asks what is wrong -- in Twi, in the spoken register -- paired with a single English phrase a health worker can read at a glance to see exactly what the issue is.

symptom_twi   Nnansa ni na me nipadua ahyew, anadwo biara mefiri me nna mu fri nsuo.
tag_en        Three-day fever with night sweats
body_system   General & Fever

The English tag is the point of the dataset: it crosses the language barrier in one line, so a Twi-speaking patient's own words can be routed, triaged or searched by someone who does not read Twi.

Where the content comes from

1. 1.Source speech. `ghanaopendata/twi-health-asr-gemini-500hrs` -- Twi audio from Ghanaian health and wellness programmes, machine-transcribed. Only the transcriptions were used here; no audio was consumed.
2. 2.Grounded generation. For each transcribed passage, a model was asked to write up to 8 first-person Twi symptom descriptions, reusing the body vocabulary and turns of phrase in that passage so the wording follows how people in the corpus actually talk about illness rather than translated textbook prose. Passages with no health content -- programme idents, greetings, adverts -- returned nothing and contributed no rows.
3. 3.Tagging. Each description carries one fine-grained English triage phrase and one body_system label from a closed set of 17.
4. 4.Phonemisation. ipa_twi was produced with ghana-g2p using its native Asante Twi rules, not donor-language rules. Each word is phonemised separately and the results joined with single spaces, so ipa_twi lines up token-for-token with symptom_twi. Punctuation and digits are dropped.
5. 5.Validation. Rows were dropped unless the description read as Twi, was in the first person, and the tag was an English phrase of the right shape and specificity -- not a single generic word, not a full sentence.

Columns

Six columns, each carrying something the others do not. row_id maps back to the exact audio chunk in `ghanaopendata/twi-health-asr-gemini-500hrs` -- shard number and row offset, zero-padded -- so any row can be traced to the speech that grounded it.

body_system labels

• —General & Fever
• —Head & Neurological
• —Eyes
• —Ear, Nose & Throat
• —Respiratory & Chest
• —Heart & Circulation
• —Digestive & Abdomen
• —Urinary
• —Sexual & Reproductive Health
• —Pregnancy & Childbirth
• —Child Health
• —Skin & Wounds
• —Bones, Joints & Muscles
• —Mental & Emotional Health
• —Nutrition & Metabolic
• —Blood & Immune
• —Teeth & Mouth

Unlike a free-text category column, this vocabulary is closed and small, so it is usable as a label directly -- no clustering step needed. The fine-grained detail lives in tag_en.

What it's useful for

• —Symptom triage and intent classification in Twi -- map a patient utterance to a presenting problem, with the English tag as the target label.
• —Twi health NLP where almost no in-domain text exists: the vocabulary of complaint, body parts, duration and severity in a low-resource language.
• —Cross-lingual retrieval and routing: Twi in, English out, in one hop.
• —Prompts and evaluation sets for Twi-language health assistants, CHPS intake tools and IVR systems, where the patient's register matters as much as the content.
• —Synthesising training data for Twi medical ASR and TTS -- the descriptions are short, spoken-style utterances suited to TTS or read-speech collection, and ipa_twi gives a phoneme target without needing a G2P step at training time.

Limitations -- read before using

• —The descriptions are model-written, not collected from patients. They are grounded in real Twi health speech, but no real person said these words about their own body. This is synthetic data.
• —No clinician reviewed any of it. The tags are not verified triage labels and the descriptions are not verified clinical presentations. Do not use this for diagnosis, triage or any clinical decision-making without qualified medical review.
• —The Twi is unreviewed machine output. Grammar, tone marking and idiom will contain errors, and the register may not match every Twi-speaking community.
• —Coverage follows the source programmes. Conditions discussed on Ghanaian health television are over-represented; anything not discussed is absent. This is not a balanced picture of presentations in Ghana.
• —The source transcriptions are themselves machine-generated and carry their own errors, which can propagate into the wording here.
• —`ipa_twi` is rule-based, not verified. Phoneme units run together within a word and many are multi-character (ny, kp, gb, kʰ), so nothing marks where one phoneme ends and the next begins inside a word -- if you need unit boundaries, regenerate with a separator rather than splitting on characters. A handful of characters outside the Twi rule set were dropped rather than guessed: q (144 occurrences) and ŋ (65), which are not part of Asante Twi orthography and come from code-switching or transcription noise.

Human review by Twi speakers and by clinicians is the highest-value next step, and the row layout -- description, tag and source passage side by side -- is built for exactly that.

License

CC-BY-NC-4.0, inherited from the source dataset `ghanaopendata/twi-health-asr-gemini-500hrs`.