Bayesian Modeling to Predict Tumor Biological Parameters for Personalized Radiotherapy
Abstract
Purpose
In current clinical practice, RT parameters such as total dose and dose per fraction are not limited to but primarily determined by the TNM stage and histology of the tumor. In this standardized treatment approach, interpatient tumor heterogeneities are not accounted for, resulting in suboptimal treatment or unnecessary toxicity. To address this problem, we validated a Bayesian digital twin framework for predicting patient-specific tumor biological parameters using limited early clinical data.
Methods
A Bayesian calibration model based on the current literature was constructed to predict three key biological parameters: (i) proliferation rate, (ii) radiosensitivity, and (iii) carrying capacity. We created 5 in silico patients with different biological parameters that were not visible to the model. The mathematical foundation of the model is based on a logistic growth model coupled with a linear-quadratic term to account for killing by radiation. The priors for the model were the historical population-based data for these three biological parameters. The model was updated with longitudinal time-series data on tumor volume measurements at three specific time points: baseline (day 0), pretreatment (day 20), and after one week of treatment (day 27).
Results
For all 5 in silico patient ground truth was within the 95% confidence interval of the posterior distributions of these biological parameters, making it possible to forecast the outcome of the treatment with early clinical data.
Conclusion
Our results indicate that a Bayesian digital twin-based framework can be effectively used to predict key tumor-specific biological parameters with limited early treatment data. However, it was assumed in the study that the virtual patient tumor follows logistic growth dynamics, which may not capture the full complexity of in vivo biology. Real-world validation is therefore required to confirm the validity of this concept.