Recurrent high-grade glioma (HGG) remains incurable with inevitable evolution of resistance and high inter-patient heterogeneity in time to progression (TTP). Here, we evaluate if early tumor volume response dynamics can calibrate a mathematical model to predict patient-specific resistance to develop opportunities for treatment adaptation for patients with a high risk of progression. A total of 95 T1-weighted contrast-enhanced (T1post) MRIs from 14 patients treated in a phase I clinical trial with hypo-fractionated stereotactic radiation (HFSRT; 6 Gy × 5) plus pembrolizumab (100 or 200 mg, every 3 weeks) and bevacizumab (10 mg/kg, every 2 weeks; NCT02313272) were delineated to derive longitudinal tumor volumes. We developed, calibrated, and validated a mathematical model that simulates and forecasts tumor volume dynamics with rate of resistance evolution as the single patient-specific parameter. Model prediction performance is evaluated based on how early progression is predicted and the number of false-negative predictions. The model with one patient-specific parameter describing the rate of evolution of resistance to therapy fits untrained data ( ${\mathrm{R}}^2=0.70$ ). In a leave-one-out study, for the nine patients that had T1post tumor volumes ≥1 cm³, the model was able to predict progression on average two imaging cycles early, with a median of 9.3 (range: 3–39.3) weeks early (median progression-free survival was 27.4 weeks). Our results demonstrate that early tumor volume dynamics measured on T1post MRI has the potential to predict progression following the protocol therapy in select patients with recurrent HGG. Future work will include testing on an independent patient dataset and evaluation of the developed framework on T2/FLAIR-derived data.

Patients with recurrent high‐grade glioma (rHGG) have a poor prognosis with median progression‐free survival (PFS) of <7 months. Responses to treatment are heterogenous, suggesting a clinical need for prognostic models. Bayesian data analysis can exploit individual patient follow‐up imaging studies to adaptively predict the risk of progression. We propose a novel sample size analysis for Bayesian individual dynamic predictions and demonstrate proof of principle. We coupled a nonlinear mixed effects tumor growth inhibition model with a survival model. Longitudinal tumor volumes and time‐to‐progression were simulated for 2000 in silico rHGG patients. Bayesian individual dynamic predictions of PFS curves were evaluated using area under the receiver operating characteristic curve (AUC) and Brier skill score (BSS). We investigated the effects of sample size on AUC and BSS margins of error. A power law relationship was observed between sample size and margins of error of AUC and BSS. Sample size was also found to be negatively correlated with margins of error and landmark time. We explored the use of this sample size analysis as a clinical look‐up table for prospective clinical trial design and retrospective clinical data analysis. Here, we motivate the application of Bayesian individual dynamic predictions as a clinical end point for clinical trial design. Doing so could aid in the development of study protocols with patient‐specific adaptations (escalate or de‐escalate dose or frequency of drug administration, increase or decrease the frequency of follow‐up, or change therapeutic modality) according to patient‐specific prognosis. Future developments of this approach will focus on further model development and validation.

1 1.1 Background: Radiotherapy (RT) is a localized therapy used to treat approximately 50% of all cancer patients and 75% of head and neck cancer (HNC) patients. Despite tumor response to RT, off-target effects to surrounding organs at risk (OAR) can result in toxicities that exacerbate patient symptoms and necessitate interventions such as tube feeding. While current models of normal tissue complication probability (NTCP) can account for complex OARs-symptoms interactions, they only allow for dichotomous outcomes without consideration of graded response. Polytomous patient-reported outcomes (PROs), collected via patient surveys, capture patients’ experience of their symptoms throughout treatment. Unfortunately, such measures are rarely incorporated when modeling treatment-related toxicities. 1.2 Purpose: In this study, we aimed to develop a model of NTCP to recapitulate population-level PRO dynamics from a published dataset and explore response and cross-toxicity trade-offs in HNC. 1.3 Methods: We employed the classical linear-quadratic dose-response model to describe tumor response to RT with logistic growth. As a surrogate for normal tissue damage, we modeled absorbed dose kinetics to individual OARs (buccal mucosa, oral cavity, superior pharyngeal constrictor muscle (sPCM), and body) using a one-compartment pharmacokinetic model with linear elimination. We then employed a minimal inhomogeneous continuous-time Markov chain model (with tumor size and absorbed dose as time-varying covariates) to describe PRO dynamics (oral pain, dysphagia, weight loss, and tube feeding, measured on the EORTC-HN35 scale). We modeled symptom-specific NTCP as the cumulative incidence of severe symptom. Finally, we explored various response-toxicity and toxicity-toxicity trade-offs with respect to dose, OAR sparing, and fractionation. 1.4 Results: The developed mathematical model recapitulated the qualitative features of the motivating published dataset, including 1) a transient reduction of a subset of symptoms for 1–2 weeks; followed by 2) an acute exacerbation of symptoms throughout the rest of RT; followed by 3) a long-term relaxation of symptoms to below baseline levels. Response-toxicity trade-offs were sensitive to dose, variably sensitive to OAR sparing (dependent on OAR-symptom associations), and sensitive to fractionation. Toxicity-toxicity trade-offs were insensitive to dose, sensitive to OAR sparing, and insensitive to fractionation. 1.5 Conclusions: Overall, this mathematical model addresses some of the limitations of current NTCP models by explaining graded toxicity dynamics. By integrating both tumor control and quality of life considerations into a singular model, recommendations of dose, OAR sparing, and fractionation can be made. Future iterations of the model could aid clinicians in RT dose-finding and selecting a RT plan that will optimize tumor control and patients’ quality of life.