Summary
A preprint describes a causal digital-twin framework that estimates individual survival benefit from chemoradiotherapy versus radiotherapy alone in head and neck cancer, stratified by HPV status.
Contents
Background
Head and neck cancer treatment often requires a choice between radiotherapy alone and chemoradiotherapy (CRT). Adding chemotherapy can improve survival but also increases toxicity. Selecting patients who are most likely to benefit is difficult, and current decisions rely heavily on population-level evidence. A preprint posted on medRxiv describes a causal digital-twin framework that estimates individual-level survival benefit from CRT versus radiotherapy alone, stratified by HPV status.
The causal digital-twin framework
The model was developed by Keyur D. Shah and colleagues at Emory University using the public RADCURE cohort of 3,346 patients. Patients were divided into HPV-positive and HPV-negative groups and further split into training (radiotherapy start year ≤ 2007) and test (start year > 2007) sets. A Bayesian-network structure learning approach identified the minimal set of features causally related to 2-year survival (SVy2). A causal-forest model, based on double machine learning, then estimated the individualized treatment effect τ, defined as the difference in predicted 2-year survival probability under CRT versus radiotherapy alone. Overlap restrictions were applied to avoid extrapolation.
Key findings
In the HPV-negative group, the model showed strong discriminative performance, with a test-set AUC of 0.72 and a C-index of 0.81 (p < 0.01). The average treatment effect across the test set favoured CRT, with a mean τ̂ of 0.13, indicating a modest but positive improvement in 2-year survival. The model’s individualized estimates were clinically coherent: patients with node-positive disease or good performance status tended to show greater benefit, while those with high tumour burden or poor performance status showed less. Decision-curve analysis indicated that a treatment strategy guided by the model’s τ̂ estimates yielded higher net benefit than uniform treatment policies.
Among HPV-positive patients, treatment-effect heterogeneity was attenuated, consistent with the known more favourable prognosis of this subgroup. The model therefore captured differences in the potential value of adding chemotherapy across HPV strata.
Implications and limitations
The authors describe this as a proof-of-concept for an interpretable, patient-specific survival simulation tool in head and neck cancer. The framework could eventually support individualized treatment decision-making by quantifying the expected benefit of CRT for a given patient. However, the work is limited to internal evaluation using a single retrospective cohort, and external validation in independent datasets is necessary. The model currently estimates 2-year survival only, and its utility in prospective clinical settings remains untested. The preprint has not yet undergone peer review.