OA Stage-Timed Control — Real Data Test Pack

This pack gives you working Python code plus one-command scripts to download real, openly licensed datasets from the internet and regenerate figures.

What's inside

• scripts/fetch_real_data.sh / fetch_real_data.ps1 — download real CSVs:
  • Toe-in gait dataset (ADAMTS/MMP surrogate: KAM reduction with toe-in) — GitHub / Apache-2.0
  • UCI Multivariate Gait Data (gait.csv, CC BY 4.0)
• src/ — minimal, auditable pipeline:
  • loads real CSVs if present, else falls back to tiny built-in demo
  • trains a light classifier (logistic) to trigger stage-switches
  • computes ROC, PR, calibration, decision curve net benefit
  • exports 8 figures in figures/
• figures/ — outputs go here
• data/ — put CSVs here (scripts will download them)

Note: I cannot directly bundle third‑party files from the web in this environment. Run the fetch script below on your machine to pull the real CSVs into data/.

Quick start

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
bash scripts/fetch_real_data.sh     # downloads real CSVs into data/
python src/run_pipeline.py          # builds dataset & saves 8 figures

Windows PowerShell:

py -m venv .venv; .\.venv\Scripts\Activate.ps1
pip install -r requirements.txt
powershell -ExecutionPolicy Bypass -File scripts\fetch_real_data.ps1
python src\run_pipeline.py

Real datasets used

• ToeInKAMReduction (X_TIdiff.csv, y_TIP1diff.csv), Suhlrich et al. — Apache-2.0 (GitHub).
• Multivariate Gait Data (gait.csv) — UCI Machine Learning Repository, CC BY 4.0.

See LICENSE-3RD-PARTY.md for sources and license notes.