Forecasting Individual Patient Response to Radiotherapy with a Dynamic Carrying Capacity Model in Head and Neck Cancer

Nearly 66% of all cancer patients receive radiotherapy (RT). Currently, RT scheduling does not take into consideration tumor volume dynamics. If response to an RT schedule can be predicted accurately, then there is a potential for treatment adjustment. The objectives of this study are to model tumor volume dynamics in response to RT and to evaluate the patient-specific predictive power of the model for patient outcomes. Tumor volume data were collected for 2 independent cohorts of head and neck cancer patients from Moffitt Cancer Center (MCC) and M.D. Anderson Cancer Center (MDACC) that received 66-70 Gy RT in 2 Gy daily fractions. Tumor volume measurements were derived from CT scans: 2 before RT and weekly scans during RT. Tumor growth was described with a logistic growth model with intrinsic growth rate, λ, and tumor carrying capacity, K. The effect of RT was modeled as an instantaneous reduction in carrying capacity with fraction δ. To predict response to RT for individual patients, we combined the distribution of MCC-learned δ values and weekly measurements of volume reduction in the untrained MDACC cohort to estimate δ to predict volume reduction and patient outcomes. The model fit data from MCC with patient-specific values for λ and δ with high accuracy (R2 = 0.95). Model analysis revealed that growth rate λ is not patient specific. A uniform λ reduces R2 to 0.92 while reducing the number of free parameters in the model (K and δ being patient specific). This MCC-trained model was then cross-validated on the independent cohort from MDACC (R2 = 0.98), demonstrating transferability of λ. The trained model predicts patient-specific RT responses with >70% accuracy for loco-regional control and disease-free survival without considering any patient-specific observations, and inclusion of on-treatment observations further increases prediction accuracy.