310 - Digital Twins of Clinical Trials in Head and Neck Squamous Cell Carcinoma: In Silico Models of Chemoradiation Resistance and Recurrence Patterns Predict Clinical Outcomes
Presenter(s)
J. B. Stevens1, Y. M. Mowery2, J. Smith1, C. Wang1, J. G. Liu1, D. M. Brizel3, and K. Lafata1; 1Duke University, Durham, NC, 2Department of Radiation Oncology, UPMC Hillman Cancer Center, Pittsburgh, PA, 3Department of Head and Neck Surgery and Communication Sciences, Duke University Medical Center, Durham, NC
Purpose/Objective(s): In silico AI models may enhance understanding of factors driving therapeutic response for patients with head and neck squamous cell carcinoma (HNSCC). We hypothesized that computational models of chemoradiation response, parameterized with 18F-FDG-PET radiomics and calibrated using intra-treatment imaging, can predict progression-free survival (PFS) and guide future clinical trial design.
Materials/Methods: Change in tumor volume was simulated using an ordinary differential equation with Gompertz growth dynamics and a hybrid log-cell kill/linear-quadratic chemoradiation model. Simulations were initialized using pre-treatment 18F-FDG-PET/CT images of patients undergoing chemoradiation therapy (70 Gy) enrolled on a single-institution prospective clinical trial (NCT01908504), where intra-treatment (20 Gy) imaging was used to evaluate early metabolic response. Radiomic features SUVmean and MTV50 were extracted from primary and nodal GTVs and used to initialize the model. Treatment sensitivity parameters, stratified by HPV status, were calibrated by matching simulated MTV50 data to observed intra-treatment MTV50 data via Wasserstein distance minimization. Mann-Whitney U-test was used to evaluate the resulting difference between calibrated and observed MTV50 data. Recurrence thresholds were optimized by fitting simulated PFS to RTOG 0129 data using a negative log-likelihood loss function. To evaluate the fit, 3-year PFS was calculated via Kaplan-Meier analysis. The model was validated by comparing predicted PFS to NCT01908504 results. Uncertainty was quantified by simulating 100 digital twin trials and prediction accuracy was evaluated using integrated brier scores (IBS).Results: The prospective cohort comprised 87 HNSCC patients (n=55 HPV+ oropharyngeal carcinoma, median follow-up 47.6 months). Calibration with intra-treatment MTV50 yielded good agreement between simulated and observed distributions (HPV+ median 2.00 cm3 vs 1.85 cm3, p=0.87; HPV- median 1.94 cm3 vs 2.02 cm3, p=0.98). Optimization of recurrence thresholds demonstrated strong agreement (p=0.32, log-rank) of 3-year PFS between simulated (0.631, 95% CI, 0.610—0.652) and RTOG 0129 (0.628, 95% CI, 0.575—0.680) results. On model validation, mean IBS was 0.18 across the 100 in silico trials with variance < 0.01%, demonstrating strong predictive accuracy. Simulation predicted mean 3-year PFS 0.71 +/- 0.06 consistent with 3-year PFS of NCT01908504 (0.75, 95% CI, 0.66—0.84).
Conclusion: Digital twins of clinical trials calibrated via intra-treatment imaging predicted tumor response to chemoradiation and captured observed PFS in HPV-stratified HNSCC. This approach may inform future clinical trial design, where personalized treatment response can drive therapeutic adaptation and improve patient outcomes.