Adaptive Cold Atmospheric Plasma Cancer Treatment With Real-Time Electrochemical Impedance Spectroscopy

Abstract

Cold atmospheric plasma (CAP) jet is an ionized gas with a rich combination of reactive oxygen and nitrogen species, charged particles, and photons. It has been demonstrated the CAP is capable of eliminating cancer cells without damaging normal cells both in vitro and in vivo. However, obtaining a safe and reliable CAP treatment is challenging since the therapeutic effectiveness of the CAP treatment is affected not only by the CAP-generating parameters, such as discharge voltage and excited gas composition, but also by the surrounding environmental conditions, such as ambient temperature and humidity. Furthermore, even under the same treatment conditions, cancer cells may response differently due to their stochastic nature. To address these challenges, this dissertation presents an adaptive cold atmospheric plasma (CAP) cancer treatment framework with real-time electrochemical impedance spectroscopy (EIS) in-situ diagnostics.

First, the reinforcement learning framework is constructed. To provide the initial treatment instruction, a CAP cancer cell treatment empirical model is generalized through a Gaussian process (GP). Then the CAP cancer treatment is formulated as a Markov decision process (MDP), to which Q-learning is applied, followed by the evaluation of the RL agent through a simulated perturbed actual cell dynamics. To address the safety concerns introduced by the RL exploration-exploitation dilemma, the risk is formulated by accounting for uncertainties in a probabilistic sense. With more data becoming available through CAP treatments, the safety concerns diminish, and more aggressive treatments can be planned.

Second, the EIS-based in-situ diagnostics is studied with the canonical correlation analysis (CCA), where the correlation between impedance measurements and CAP treatment effectiveness, i.e., cell viability, is investigated. To acquire data to establish the correlation, the experiment equipment and hardware setup are developed, followed by the design of experiment conditions and procedures. Then, a Butterworth low-pass filter is designed to smooth out the noise presented in the impedance signal so that a more reliable correlation can be obtained. With processed impedance and cell viability data, the CCA is applied with regularization to improve the generalization capability. To further predict the cell viability with given impedance measurements, a linear regression (LR) is applied to the regularized CCA (RCCA) data.

Finally, with the above model for real-time diagnostics, adaptive CAP treatments are performed. For a given desired target cell viability, an adaptive experiment group is designed, at which the treatment time is determined by the real-time in-situ diagnostics with the trained RCCA and LR models. For comparison, with the same desired target, another experiment group is designed to have a fixed treatment time obtained through empirical data prior to the CAP treatments. When the treatment conditions are consistent with the conditions utilized to extract CCA correlations, namely the nominal conditions, it is demonstrated that the RCCA and LR models can adaptively choose appropriate treatment time for the adaptive group. To further evaluate the potent of the trained models, when the treatment conditions are varied, i.e., perturbed conditions, the adaptive experiments are conducted in the same manner. As a result, while the treatments in the fixed group have mismatched effectiveness, the treatments in the adaptive group can still adjust their treatment time to reach to the desired target cell viability, by utilizing the real-time impedance measurements to detect the discrepancy between the empirical knowledge and the actual cell responses.