The Medkit-Lear(ig) Eviromet

Alex J. Cha, Ioaa Bica, Aliha Huyuk, Daiel Jarrett, ad Mihaela va der Schaar

The Medkit-Lear(ig) Eviromet, or Medkit, is a publicly available Pytho package providig simple ad easy access to high-fidelity sythetic medical data.

Primarily, Medkit is a tool that supports: (1) a variety of realistic eviromet models—leared from actual data, to reflect real medical settigs), thus allowig simulatio of (2) a variety of expressive ad customisable policy models that represet complex huma decisio-behaviours; as well as (3) esurig that the eviromet ad policy compoets are disetagled—hece idepedetly cotrollable.

By fulfillig the above, Medkit seeks to eable advaces i decisio modellig to be validated more easily ad robustly by eablig users to obtai batch datasets with kow groud-truth policy parameterisatios that simulate decisio makig behaviours with various degrees of Markoviaity, bouded ratioality, cofoudig, idividual cosistecy ad variatio i practice.

Medkit is pip istallable - to work with the latest versio, we recommed cloig it, optioally creatig a virtual ev, ad istallig it (this will automatically istall depedecies):

git cloe https://github.com/XaderJC/medkit-lear.git

cd medkit-lear

pip istall -e .

Alteratively, Medkit is available o PyPI, ad ca be istalled simply with:

pip istall medkit-lear

Example usage:

import medkit as mk

sythetic_dataset = mk.batch_geerate(
    domai = "Ward",
    eviromet = "CRN",
    policy = "LSTM",
    size = 1000,
    test_size = 200,
    max_legth = 10,
    scale = True
)

static_trai, observatios_trai, actios_trai = sythetic_dataset['traiig']
static_test,  observatios_test,  actios_test  = sythetic_dataset['testig']

While medical machie learig is by ecessity almost always etirely offlie, we also provide a iterface through which you ca iteract olie with the eviromet should you fid that useful. For example, you could trai a custom RL policy o this eviromet with a specified reward fuctio, the you ca test iferece algorithms o their ability to represet the policy.

ev = mk.live_simulate(
    domai="ICU",
    eviromet="SVAE"
)

static_obs, observatio, ifo = ev.reset()
observatio, reward, ifo, doe = ev.step(actio)

Citig

If you use this software please cite as follows:

@iproceedigs{cha2021medkitlear,
        title={The Medkit-Lear(ig) Eviromet: Medical Decisio Modellig through Simulatio},
        author={Alex James Cha ad Ioaa Bica ad Aliha H{\"u}y{\"u}k ad Daiel Jarrett ad Mihaela va der Schaar},
        booktitle={Proceedigs of the Neural Iformatio Processig Systems Track o Datasets ad Bechmarks},
        year={2021},
        url={https://opereview.et/forum?id=Ayf90B1yESX}
}

The Medkit-Learn(ing) Environment Alex J. Chan, Ioana Bica, Alihan Huyuk, Daniel Jarrett, and Mihael