Cancer: Getting Started#
Research question#
How can clinical features and pathology patch features be combined to explore time-to-event prediction? The starter example trains an intermediate-fusion Cox model using a small synthetic cohort.
No real patient data are needed. Real-data HANCOCK examples require separate dataset access and additional setup.
Prepare the repository#
Use the KaleCancer repository, Python 3.11 or 3.12, and uv. From the repository root, follow its quickstart:
uv sync --all-extras
The public repository contains the example paths below. Confirm the session environment with the organisers before installing; these documentation checks do not replace testing the dependencies on your machine.
Run the synthetic example#
uv run python examples/synthetic/synthetic_survival.py
This CPU-capable, no-download demonstration generates clinical variables, pathology patch features and censored survival times, then trains the model.
The repository also provides a separate synthetic survival example notebook. The session’s Live-Coding Notebook is available here with its saved outputs and a Colab link.
What to look for#
How the synthetic cohort is assembled.
Which clinical and patch inputs are used.
Where intermediate fusion occurs.
Which evaluation metrics are printed.
Continue the track#
Hands-on Tutorial: locate the synthetic inputs and identify the prediction task.
Live-Coding Notebook: follow data loading, modelling, training and evaluation.
Results and Interpretation: understand the output and limits of a synthetic survival exercise.
Technical references#
Within the software repository, consult README.md, docs/quickstart.md, examples/synthetic/synthetic_survival.py, and examples/hancock/.
For library support, use PyKale discussions or the repository’s issue tracker.
Important
Synthetic scores are pipeline checks, not evidence of clinical predictive performance. Clinical data require appropriate permissions and secure infrastructure.