Reinforcement Learning With Safe Exploration for Adaptive Plasma Cancer Treatment

Cold atmospheric plasma (CAP) jet is an ionized gas with a rich combination of reactive oxygen/nitrogen species, charged particles, and photons. By both
in vitro
and
in vivo
experiments, it has been demonstrated that CAP can be successfully utilized in cancer treatments. However, the therapeutic effects of CAP depend on various factors, such as plasma discharge voltage, gas composition, treatment duration, and type of cancer cells. This article presents an adaptive plasma system, where the CAP treatment conditions are adjusted online depending on the actual cancer cell response. In particular, we utilize safe
$Q$
-learning to schedule CAP cancer treatment autonomously while prohibiting excessive treatments caused by large uncertainties. As it is capable of learning the dynamic characteristics of the particular cancer cell under treatments
in situ
, we can treat cancer successfully without the complete prior knowledge of cancer characteristics, where the uncertainties of the learned dynamics are carefully accounted. The efficacy of the proposed algorithm is illustrated by numerical examples with an empirical cancer dynamic model constructed from
in vitro
experiments.