Autoencoder-Based Architecture for Parameter Estimation of a Tumor Model

Chemotherapy remains one of the predominant treatment modalities for cancer patients. However, current drug dosages and schedules typically do not account for intra-and interindividual pharmacokinetic and pharmacodynamic variations. Our approach aims to formulate individual patient characteristics as mathematical models. By determining the parameters of these models, we can identify patient-specific differences and establish optimal, personalized therapy in an automated manner. In this study, we determine the parameters of a previously validated system of differential equations that describe tumor dynamics and pharmacokinetics. We have developed a novel autoencoder-based architecture for identifying the parameters of this model. Utilizing time series data of tumor volumes and administered doses, this architecture effectively estimates the model parameters. We trained and tested the network with three distinct input lengths, and the results demonstrated that the incorporation of the tumor model can increase the accuracy and accelerate the training. The proposed algorithm shows potential for future applications in identifying model parameters, which could be utilized in the design of therapeutic regimens.