3614 - A Causal Digital Twin Framework for Personalized Chemoradiotherapy Intensification in Head and Neck Cancer

Purpose/Objective(s): Treatment selection in head and neck cancer (HNC) requires balancing potential survival benefit from concurrent chemoradiotherapy (CRT) against treatment-related toxicity. We hypothesized that a causal digital twin (DT) framework could estimate individualized counterfactual 2-year survival (SVy2) under CRT versus radiotherapy (RT) alone (treatment intensification defined as addition of concurrent chemotherapy), quantify heterogeneity of treatment effect, and generalize across independent cohorts.

Materials/Methods: A causal forest-based DT model was developed and trained on a temporally split development cohort (n=2174; 41% CRT; 78% SVy2; 29.3% HPV-positive) from a publicly available single-institution dataset using harmonized key clinical predictors (primary site, T stage, gross tumor volume, HPV status) identified via Markov blanket feature selection. SVy2 was defined as survival at or beyond 2 years from RT initiation. The DT model was used to estimate individual treatment effects (ITE), meaning the expected difference in outcome for each patient if treated with CRT versus RT alone, in a temporally held-out internal test cohort (n=820; 43% CRT; 53% SVy2; 38% HPV-positive). Performance was further evaluated in two external validation cohorts (combined n=307; 64% CRT; 84% SVy2; 25% HPV-positive), including one multi-institutional dataset. To prevent biased comparisons, patients whose clinical characteristics made them almost certain to receive/not receive CRT (treatment propensity <0.1 or >0.9) were excluded. A doubly robust estimation approach was applied to reduce bias if either the outcome model or the treatment model was imperfect. Finally, differences in treatment benefit across patient subgroups were assessed by HPV phenotype to evaluate treatment effect heterogeneity.

Results: In the internal test cohort, the average estimated improvement in SVy2 with CRT (the mean ITE) was 8.2% (95% CI 7.4-9.0%). In external validation, the average estimated improvement was 13.3% (95% CI 12.4-14.1%). Estimated treatment benefit was seen in both HPV-positive and HPV-negative groups in the internal cohort (11.2% and 10.0%) and in the external cohorts (16.2% and 13.9%), respectively. More than 95% of patients met the necessary overlap between treatment groups to allow valid comparison of outcomes under alternative treatment scenarios. Estimates generated using a secondary “doubly robust” method showed the same overall direction of effect as the primary causal forest results.

Conclusion: A causal DT framework can estimate, for each individual patient, how 2-year survival would differ with CRT versus RT alone in HNC, and shows consistent evidence of added benefit from chemotherapy intensification across independent cohorts and HPV-positive and HPV-negative groups. By leveraging familiar clinical variables, this interpretable, patient-level simulation tool supports personalized treatment decisions and motivates prospective evaluation.