NeuroRisk — Voice Memory Screening on MCP + Amazon SageMaker

Voice-based early cognitive screening for neurodegenerative disease risk, built as a self-hosted MCP server (spec 2025-11-25, Streamable HTTP) with a simulated Alexa+ voice experience and a machine-learning risk model deployed on Amazon SageMaker.

Demo screening trained on synthetic data. Not a medical diagnosis.

What it does

A user opens the web simulator, taps the microphone (or types), and the Alexa+-style voice assistant walks them through an 8-step cognitive screening adapted from validated brief tests:

1. Age collection

2. Three-word delayed recall (memory_intro)

3. Orientation in time (day, month, year)

4. Forgetting scheduled events (self-report)

5. Usual location of everyday objects

6. Misplacing items at home (self-report)

7. Verbal fluency — name as many animals as possible in 60 seconds

8. Delayed recall of the three words

The answers are scored and sent to an Amazon SageMaker endpoint that returns a risk probability (0–1). The result, a risk band (low / intermediate / high), and a plain-language message are shown immediately. All results are stored in Amazon S3 so users can track their progress over multiple sessions.

Related MCP server: nirog-mcp

Architecture

Browser (simulator.html)
        │  Streamable HTTP  (MCP 2025-11-25)
        ▼
server.py  ── FastMCP (mcp[cli])
        │
        ├── boto3 ──► Amazon SageMaker  (cognitive-risk endpoint)
        │                  XGBoost model, binary classification
        └── boto3 ──► Amazon S3  (neurorisk-hackathon-henrique-2026)
                           └── output/history/<user_id>/*.json

The simulator is served by the MCP server itself at /, so the browser and the MCP endpoint share the same origin — no CORS configuration needed.

MCP tools exposed

AWS services used

Requirements

• Python 3.11+

• An AWS account with the following resources configured:

  • SageMaker endpoint — a binary classification model deployed and active (XGBoost recommended). The default expected name is cognitive-risk, but you can override it.

  • S3 bucket — used to store screening history per user. Create a bucket in the same region as your endpoint.

  • IAM permissions — your credentials must allow sagemaker:InvokeEndpoint, s3:PutObject, s3:GetObject, and s3:ListObjectsV2.

• AWS credentials configured locally via aws configure or environment variables.

pip install -r requirements.txt

Configuration

Before running, set the following environment variables to point to your own AWS resources:

You can set them inline, or create a .env file and export them before running.

AWS credentials are also picked up automatically from ~/.aws/credentials if you have the AWS CLI configured (aws configure).

Running

1. Install dependencies

pip install -r requirements.txt

2. Set your AWS environment variables

export AWS_DEFAULT_REGION=us-east-1        # your region
export AWS_ACCESS_KEY_ID=AKIA...           # your key
export AWS_SECRET_ACCESS_KEY=...           # your secret
export BUCKET=my-bucket-name               # your S3 bucket
export ENDPOINT_NAME=my-endpoint-name      # your SageMaker endpoint

3. Start the server

python server.py

4. Open the simulator

Open http://localhost:8000 in Chrome or Edge (speech recognition requires a Chromium-based browser).

The MCP endpoint is available at http://localhost:8000/mcp.

Testing

python test_client.py

Runs a full MCP handshake and calls all four tools against the running server.

Hackathon track

• Primary: Alexa+ (self-hosted MCP server, spec 2025-11-25, Streamable HTTP)

• Mini challenge: AWS Builder (SageMaker + S3)

License

MIT — see LICENSE.