Speaker
Description
Graves' disease is an autoimmune thyroid disorder causing hyperthyroidism, with a lifetime risk of 3% in women and 0.5% in men, see \cite{A}. It is commonly treated with antithyroid drugs where current dosing guidelines still provide limited support for optimal dose selection as it is reported e.g. in \cite{B,C}. Strong inter-individual variability poses a major challenge for conventional computer-based dosing approaches. Among machine learning paradigms, reinforcement learning is particularly well suited for managing sequential treatment decisions (see e.g. \cite{D}) as required in Graves' disease. We developed a reinforcement learning based agent, which approximates its policy using a neural network. It was evaluated on a validated simulation platform and outperformed existing algorithmic approaches as well as experienced endocrinologists. Since the decisions of neural network-based agents are opaque to clinicians, we additionally developed fully human-readable, inherently interpretable tabular reinforcement learning agents. Contrary to common expectations, these transparent agents performed better than the neural network–based approach. This result suggests that transparent, tabular reinforcement learning may be applicable to a broader class of cyclic treatment settings in which drug doses are iteratively adjusted based on measurable physiological parameters, as it is the case in Graves' disease.
Bibliography
@article{A,
title = {Graves’ Disease},
volume = {375},
ISSN = {1533-4406},
url = {http://dx.doi.org/10.1056/NEJMra1510030},
DOI = {10.1056/nejmra1510030},
number = {16},
journal = {New England Journal of Medicine},
publisher = {Massachusetts Medical Society},
author = {Smith, Terry J. and Heged\"{u}s, Laszlo},
editor = {Longo, Dan L.},
year = {2016},
month = oct,
pages = {1552–1565},
}
@article {B,
author = {George J. Kahaly and Luigi Bartalena and Lazlo Hegedüs and Laurence Leenhardt and Kris Poppe and Simon H. Pearce},
title = {2018 European Thyroid Association Guideline for the Management of Graves’ Hyperthyroidism},
journal = {European Thyroid Journal},
year = {2018},
publisher = {S. Karger AG},
volume = {7},
number = {4},
pages= {167 - 186},
}
@article{C,
title = {Hyperthyroidism therapy: What can decision support systems already achieve?},
volume = {16},
ISSN = {1998-7781},
number = {4},
journal = {Journal f\"{u}r Klinische Endokrinologie und Stoffwechsel},
publisher = {Springer Science and Business Media LLC},
author = {Benninger, Thomas and Theiler-Schwetz, Verena and Pilz, Stefan and Trummer, Christian and Reichhartinger, Markus},
year = {2023},
pages = {122–131},
}
@article{D,
author = {Jayaraman, Pushkala and Desman, Jacob and Sabounchi, Moein and Nadkarni, Girish N and Sakhuja, Ankit},
journal = {NPJ Digit. Med.},
title = {A primer on reinforcement learning in medicine for clinicians},
year = {2024},
month = nov,
number = {1},
pages = {337},
volume = {7},
publisher = {Springer Science and Business Media LLC},
}
