Datasaur 2026 | Qazcode Challenge

Medical Diagnosis Assistant: Symptoms → ICD-10

An AI-powered clinical decision support system that converts patient symptoms into structured diagnoses with ICD-10 codes, built on Kazakhstan clinical protocols.

---

Challenge Overview

Participants will build an MVP product where users input symptoms as free text and receive:

• Top-N probable diagnoses ranked by likelihood
• ICD-10 codes for each diagnosis
• Brief clinical explanations based on official Kazakhstan protocols

The solution must run using GPT-OSS — no external LLM API calls allowed. Refer to notebooks/llm_api_examples.ipynb

---

Data Sources

Kazakhstan Clinical Protocols

Official clinical guidelines serving as the primary knowledge base for diagnoses and diagnostic criteria.[corpus.zip]

Data Format

{"protocol_id": "p_d57148b2d4", "source_file": "HELLP-СИНДРОМ.pdf", "title": "Одобрен", "icd_codes": ["O00", "O99"], "text": "Одобрен Объединенной комиссией по качеству медицинских услуг Министерства здравоохранения Республики Казахстан от «13» января 2023 года Протокол №177 КЛИНИЧЕСКИЙ ПРОТОКОЛ ДИАГНОСТИКИ И ЛЕЧЕНИЯ HELLP-СИНДРОМ I. ВВОДНАЯ ЧАСТЬ 1.1 Код(ы) МКБ-10: Код МКБ-10 O00-O99 Беременность, роды и послеродовой период О14.2 HELLP-синдром 1.2 Дата разработки/пересмотра протокола: 2022 год. ..."}

---

Evaluation

Metrics

• Primary metrics: Accuracy@1, Recall@3, Latency
• Test set:: Dataset with cases (data/test_set), use query and gt fields.
• Holdout set: Private test cases (not included in this repository)

Product Evaluation

Working demo interface: user inputs symptoms → system returns diagnoses with ICD-10 codes;

---

Getting Started

1. Clone the repository

git clone https://github.com/dair-mus/qazcode-nu.git
cd qazcode-nu

2. Set up the environment

We kindly ask you to use uv as your Python package manager.

Make sure that uv is installed. Refer to uv documentation

uv venv
source .venv/bin/activate
uv sync

3. Running validation

You can use src/mock_server.py as an example service. (however, it has no web UI, only an endpoint for eval).

uv run uvicorn src.mock_server:app --host 127.0.0.1 --port 8000

Then run the validation pipeline in a separate terminal:

uv run python evaluate.py -e http://127.0.0.1:8000/diagnose -d ./data/test_set -n <your_team_name>

-e: endpoint (POST request) that will accept the symptoms

-d: path to the directory with protocols

-n: name of your team (please avoid special symbols)

By default, the evalutaion results will be output to data/evals.

Docker

We prepared a Dockerfile to run our mock server example.

docker build -t mock-server .
docker run -p 8000:8000 mock-server

Then run the validation as shown above.

Feel free to use the mock-server FastAPI template and Dockerfile structure to build your own project around.

Remember to adjust the CMD in Dockerfile for your real Python server instead of src.mock_server:app before submission.

Submission Checklist

• Everything packed into a single project (application, models, vector DB, indexes)
• Image builds successfully: docker build -t submission .
• Container starts and serves on port 8080: docker run -p 8080:8080 submission
• Web UI accepts free-text symptoms input
• Endpoint for POST requests accepts free-text symptoms
• Returns top-N diagnoses with ICD-10 codes
• No external network calls during inference
• README with build and run instructions

How to Submit

1. Provide a Git repository with Dockerfile
2. Submit the link via submission form
3. We will pull, build, and run your container on the private holdout set

---

Repo structure

• data/evals: evaluation results directory
• data/examples/response.json: example of a JSON response from your project endpoint
• data/test_set: use these to evaluate your solution.
• notebooks/llm_api_examples.ipynb: shows how to make a request to GPT-OSS.
• src/: solution source code would go here, has a mock_server.py as an entrypoint example.
• evaluate.py: runs the given dataset through the provided endpoint.
• pyproject.toml: describes dependencies of the project.
• uv.lock: stores the exact dependency versions, autogenerated by uv.
• Dockerfile: contains build instructions for a Docker image.