Multiprocess Dynamic Modeling of Tumor Evolution with Bayesian Tumor-Specific Predictions

We propose a sequential probabilistic mixture model for individualized tumor growth forecasting. In contrast to conventional deterministic methods for estimation and prediction of tumor evolution, we utilize all available tumor-specific observations up to the present time to approximate the unknown multi-scale process of tumor growth over time, in a stochastic context. The suggested mixture model uses prior information obtained from the general population and becomes more individualized as more observations from the tumor are sequentially taken into account. Inference can be carried out using the full, possibly multimodal, posterior, and predictive distributions instead of point estimates. In our simulation study we illustrate the superiority of the suggested multi-process dynamic linear model compared to the single process alternative. The validation of our approach was performed with experimental data from mice. The methodology suggested in the present study may provide a starting point for personalized adaptive treatment strategies.