S192
Clinical - Biomarkers of clinical response
ESTRO 2026
translation of biomarker insights into practice. This study proposes a Bayesian response-adaptive trial design that combines population-level Bayesian inference with patient-level temporal Bayesian optimization. The framework is demonstrated in the context of rectal cancer organ preservation. Material/Methods: A prospective Bayesian trial concept was developed, integrating real-time biomarker information into adaptive decision-making. The design incorporates the principle of optimal stopping and uses the image- derived Early Regression Index (ERI) as an example biomarker for guiding dose adaptation. This index was demonstrated to predict with high accuracy the probability of pathological complete remission (pCR) in rectal cancer patients receiving neoadjuvant chemoradiotherapy, based on MRI-based GTV volumetry assessed before and at mid-radiotherapy. A fully simulated clinical trial was conducted to assess feasibility, operating characteristics, and statistical performance in rectal cancer radiotherapy. Results: Simulations modeled four response classes characterized by biomarker-dependent tumor control and biomarker-independent toxicity profiles. The Bayesian adaptive algorithm efficiently identified subgroups with distinct benefit–risk profiles. An intermediate subgroup achieved improved tumor control through moderate dose escalation, while poor and excellent responders gained minimal additional benefit within acceptable toxicity limits. The proposed design achieved conclusive posterior probabilities with approximately 100 simulated patients (see Fig. 1) – substantially fewer than required by traditional randomized trial designs.
Digital Poster 3128
Histology-based prediction of distant metastasis in nasopharyngeal carcinoma via multitask deep learning Liuling Wang, Zhaodong Fei, Chuanben Chen radiation Oncology, Fujian Cancer Hospital, Fuzhou, Fujian, China Purpose/Objective: To develop and validate a multitask deep learning (DL) model based on H&E-stained whole-slide images (WSIs) for predicting distant metastasis (DM) risk and organ-specific metastatic patterns (including bone, liver, and lung) in patients with locoregionally advanced nasopharyngeal carcinoma (LANPC). Material/Methods:
This retrospective study included 147 LANPC patients. H&E-stained WSIs were preprocessed and divided into non-overlapping 224 × 224 px patches. A pathology foundation model (Uni) was used to extract 1024- dimensional features from each tile. Representative tiles were selected via unsupervised K-means clustering. A multitask attention-based multiple- instance learning framework was trained with patient- level stratification (training: n=94, validation: n=23, independent test: n=30). The model predicted overall DM risk and organ-specific metastatic patterns concurrently. Survival analysis was performed to evaluate risk stratification based on model-derived scores. Results:
Fig. 1: Median utility estimate and 95% credible intervals with varying sample sizes following the initial cohort of N = 50 patients (batches of 50 patients). Fixed 50 Gy comparator. Conclusion: This Bayesian response-adaptive design provides a quantitative framework to integrate biomarkers such as ERI into radiotherapy personalization. It enables efficient evaluation of dose–response relationships and bridges biomarker discovery with clinical translation in local cancer therapy. Keywords: Bayesian trial design, biomarker
Made with FlippingBook - Share PDF online