Protocol-Locked Danger Sign Triage Assistant for Community Health Workers in Urdu, Bangla and Arabic
Build a phone assistant that walks a community health worker through a sick child visit in Urdu, Bangla or Arabic, where the triage decision comes from an encoded clinical protocol, never from a language model. The model only does what it is good at: understanding a spoken or typed answer in a local language and turning it into a protocol field. The project tests whether that split keeps the decisions safe while making the tool easier to use than a long paper form.
Advanced
Needs a team, real planning, and a supervisor who knows the area.2 semesters · 2 to 3 students
7 to ship
4 optional extrasSuggested stack
What you should ship
- A machine-readable encoding of a published child illness triage protocol (danger signs, classifications, referral rules) with unit tests for every branch
- A phone app that runs the visit as a guided conversation in Urdu, Bangla and Arabic, typed or spoken
- A language layer that maps free answers ("he has been vomiting everything since morning") to protocol fields and asks again when unsure
- A hard rule that the classification and referral advice come only from the protocol engine
- Offline storage of visits with later sync
- A test set of scripted visits in each language, written with a clinician, and an evaluation of field extraction accuracy
- A supervisor view of visits, referrals and cases where the worker overrode the tool
If you have time left
- Voice prompts for workers with low literacy
- Pashto, Sindhi or Egyptian Arabic dialect support
- On-device extraction with a small model so the app works without a connection
- A follow-up reminder when a referred child has no recorded visit
The problem: a long protocol, a short visit
Community health workers are often the first and only health contact for families in rural Pakistan, Bangladesh, Egypt or northern Nigeria. In a sick child visit they must check a list of danger signs (not able to drink, vomiting everything, convulsions, very fast breathing, lethargy) and decide whether the child needs urgent referral. The rules are written down in clinical protocols such as the WHO Integrated Management of Childhood Illness (IMCI) guidelines, but in practice workers carry paper forms, remember part of the chart and skip steps when a household is busy.
Digital checklists help, but most are long English forms with tick boxes. A worker who thinks in Urdu, Bangla or Arabic, and talks to a mother in the same language, has to translate in her head. A chatbot sounds like the fix, but letting a language model decide whether a child is referred is unsafe: it can be fluent and wrong.
What already exists and the gap
- CommCare (Dimagi) and the Community Health Toolkit (Medic) run large health worker programmes with form-based apps, including IMCI-style workflows.
- Symptom checkers such as Ada are built for patients in a handful of languages, not for health workers following a national protocol.
- General chat assistants answer health questions in many languages, with no guarantee they follow any protocol.
The gap is an assistant that is as easy to use as a conversation in the worker's own language but whose decisions are as predictable as the paper chart.
Inspiration
The e-POCT study by Kristina Keitel and colleagues at the Swiss Tropical and Public Health Institute, published in PLOS Medicine in 2017, tested an electronic algorithm with point-of-care tests for febrile children in Tanzania against ALMANACH, an existing electronic algorithm based on IMCI, and found fewer clinical failures with far fewer antibiotic prescriptions, showing that a well-designed digital protocol can change clinical decisions. Your project keeps that idea of an explicit algorithm and adds a language layer on top.
What you build
- Health workers answer questions in their own words, by voice or text, and see a clear result: treat at home, follow up, or refer urgently, with the reason.
- Supervisors see visits, referrals and overrides, so they can coach rather than guess.
- Programme designers can update the protocol file without retraining any model.
The technically interesting part
- Separation of duties. The protocol engine is plain code with tests. The language model returns structured fields only, validated against a schema. If it is not confident, the app asks a direct yes or no question instead of guessing.
- Three scripts, three languages. Urdu and Arabic are right to left, Bangla has its own script, and real answers mix in English words. Test extraction on code-mixed answers, not only clean sentences.
- Offline first. Many visits happen with no signal. Decide which parts must work offline (the protocol engine always must) and which can wait for sync.
How to evaluate it
- Protocol correctness: every branch of the encoded protocol covered by tests, checked against the published chart by a clinician.
- Field extraction accuracy per language on your scripted visit set, with false negatives on danger signs reported separately because they matter most.
- Safety test: feed misleading or ambiguous answers and confirm the app asks again rather than downgrading a child.
- Usability: time per visit and errors compared with a paper form, in a small role-play study with nursing or public health students.
Suggested plan
- Weeks 1 to 4: study the protocol with a clinician, encode it and write tests.
- Weeks 5 to 10: the phone app and offline storage.
- Weeks 11 to 18: the language layer for one language, then the other two.
- Weeks 19 to 24: scripted visit set, evaluation and supervisor view.
- Weeks 25 to 28: role-play study, report and demo.
Path to a product
Health NGOs and government programmes already pay for health worker apps. A multilingual conversational layer that plugs into an existing protocol, or into platforms like CommCare, is a credible offering. The design pattern (language model for input, rules for decisions) also transfers to other regulated fields, which makes this a strong portfolio piece for AI engineering roles.
Scope warning
This is a research prototype. Do not use it on real patients, and do not collect real patient data without ethics approval and a partner organisation. Never let the language model produce or alter a clinical classification. Work with a doctor or nurse from the first week.
Learn what you need on Sefism
- AI engineering roadmap: structured outputs, validation and evaluation of language model features.
- Choosing LLMs roadmap: picking a model that handles Urdu, Bangla and Arabic well enough.
- AI evals cheat sheet: building the scripted visit test set.
- Mobile roadmap: offline storage and sync.
- Hugging Face Transformers documentation: running open speech and language models.
- Speech and Language Processing: the background on extraction and speech recognition.
- How do I choose which LLM to use for my project?: choosing a model you can test against a fixed protocol.
Get help with this idea
The risk is a demo that looks clever while hiding unsafe decisions. Tauseef can help you draw the line between the model and the protocol engine and design an honest safety evaluation. If you are new here, create a free account first.
Ideas and guidance, not finished projects
These are project ideas and scoping guidance, published free for students to use as a starting point. I do not build, write, or sell final-year projects, and I do not complete coursework for anyone. Take an idea, make it yours, and build it.
Not sure this one fits you?
I help students pick an idea that matches their skills and their deadline, then scope it down until it is finishable. Guidance only, never done for you.
Stuck on something specific?
Writing only gets you so far. If you want an answer to your situation rather than the general case, book a session and we will work through it together. Sessions are free for approved Sefism members, and a few slots open each week.
Follow along
New writing, resources and project ideas land here first.