What is ml4c3?

Machine Learning for Cardiology and Critical Care (ml4c3) is a pipeline for working with complex physiological data. It does several key things:

• Ingest raw data into a pre-processed format.
• Tensorize ingested data into standard hd5 or CSV formats compatible with the pipeline.
• Map desired data from hd5 or CSV files to code via the TensorMap abstraction.
• Visualize ECGs, ICU waveforms, labs, and modeling results.
• Explore summary statistics and trends.
• Train supervised and reinforcement learning models that are constructed from powerful, simple, and expressive command line arguments and TensorMaps.
• Cluster data to reveal patterns via unsupervised learning.

How do I install ml4c3?

1. Install docker and anaconda

2. Add user to group to run docker

   sudo usermod -aG docker $USER

3. Build GPU docker image:

   ./docker/build.sh

   and also the CPU-only docker image:

   ./docker/build.sh -c

4. Setup conda environment for pre-commit:

   make setup

   On macOS, you can install gmake to call gmake setup

5. Activate conda environment so that pre-commit hooks run:

   conda activate ml4c3

Why does ml4c3 exist?

When a researcher joins an academic research group, they get access to a workstation or server and some data. Sometimes a more senior researcher shares some code via email or Dropbox -- usually a long main.py script with now-outdated dependencies. Nothing is documented. A manuscript published using the code does not link to an open-source GitHub repo, nor does it sufficiently document the steps and settings needed to reproduce the results. The paper merely includes the familar phrase "code is available upon request".

The researcher spends their first week trying to set up their compute environment so they can simply run their predecessor's pipeline. The packages need updating, and a library was compiled in a special way, but nobody knows exactly how. The next two weeks are spent understanding the code. Eventually, the researcher refactors the entire pipeline. A month later, they start training models.

This story is too common in academia. We think there is a better way, so we built ml4c3 to:

• address limitations of Jupyter notebooks, one-off scripts, and glue-code.
• enhance collaborative workflow between researchers, following best practices.
• increase efficiency via excellent documentation and modular code.

How do I use ml4c3?

Read the documentation at the ml4c3 wiki.

Where do I ask questions?

Join the discussions.

Can I contribute?

Yes, we would love your help! See CONTRIBUTING to learn how to open an issue, submit a PR, and more.

Who built ml4c3?

ml4c3 was built by the Aguirre Lab at the Center for Systems Biology, Wellman Center for Photomedicine, and Cardiovascular Research Center at the Massachusetts General Hospital.

We are a team of students, postdoctoral researchers, and engineers who work on overlapping research projects at the intersection of machine learning, computer science, cardiology, and critical care.

Since we share research needs, computational tooling, and data, the benefits of closely collaborating around a shared infrastructure were obvious.