S656
Clinical – Head & neck
ESTRO 2026
harmonization.” Phys Imaging Radiat Oncol. 2023;26:100450. Keywords: Radiomics, Locoregional recurrence, Head and neck
Digital Poster 4366 An Evidence-Based Multimodal LLM Agent for Adaptive Radiotherapy Triggering in Nasopharyngeal Carcinoma: A Feasibility Study Suman Zhang 1 , Guanqun Zhou 1 , Yanfei Liu 2 , Mengyu Hao 1 , Lecheng Jia 2 , Hua Li 2 , Ying Sun 1 1 The department of Radiation Oncology, Sun Yat-sen University Cancer Center, Guangzhou, China. 2 Combination Innovation Department, Shenzhen United Imaging Healthcare Co., Ltd., Shenzhen, China Purpose/Objective: To develop and perform a feasibility assessment of a multimodal large language model (LLM) agent for automated, evidence-based decision-making on adaptive radiotherapy (ART) trigger timing in patients with nasopharyngeal carcinoma (NPC). Material/Methods: We developed an LLM agent integrating three core capabilities:1) Platform Connectivity: Automated data acquisition from hospital electronic medical record (EMR) /picture archiving and communication system (PACS) and treatment planning systems, enabling auto-contouring on fractional FBCTs, real-time morphological monitoring, and generation of reference/synthetic ART plans for dosimetric comparison.2) On-Demand Tool Library: Selectable predictive models including Normal Tissue Complication Probability (NTCP) prediction models, Tumor Control Probability (TCP) prediction models, and an Epstein–Barr virus (EBV) DNA clearance kinetics model.3) Retrieval-Augmented Generation (RAG): An NPC-specific ART literature database, continuously updated via PubMed, providing evidence for anatomical, dosimetric, and clinical indicators during ART trigger evaluation.The Qwen3-235B-based agent was used without fine-tuning. Prompting logic was optimized using 5 retrospective NPC cases, with feasibility demonstrated on 10 independent cases. Results: The proposed workflow is illustrated in Figure 1. For each patient, the agent first acquired baseline clinical information, anatomical changes, and dosimetric changes via platform tools, then automatically retrieved relevant content from the RAG database and invoked predictive tools as needed, ultimately producing a structured output. As shown in Figure 2, a representative output included the ART trigger recommendation, a trigger confidence score (range: 0- 100), the rationale for the trigger decision, and
Of 367 patients who received primary chemoradiation, 176 patients were included in the final analysis. Primary tumor sites comprised of larynx (n=79, 44%), hypopharynx (n=36, 20%), oropharynx (n=20, 11%), tonsil (n=16, 9%), oral cavity subsites (n=15, 8%), base of tongue (n=6, 3%), and other locations (n=4, 4%). 56 patients had locoregional recurrence and 120 remained disease-free at 12 months follow-up post- treatment completion. The training cohort included 141 and the hold-out test included 35 patients. The clinical-radiomics model (Naïve-Bayes with Genetic Algorithm) achieved the best test AUC of 0.82 [95% CI: 0.66-0.94], exceeding the radiomics performance of 0.78 [0.61-0.91] (Figure 1). The signature included 2 clinical and 8 radiomics features, respectively.
Conclusion: We developed a clinical-radiomics signature for LRR of locally advanced HNSCC using a prospective dataset for an Indian cohort. Subsequent prospective validation will evaluate the generalizability of this signature. References: 1. Vallières M, et al. “Radiomics strategies for risk assessment of tumour failure in head-and-neck cancer.” Sci Rep. 2017;7(1):10117. 2. Varghese AJ et al. “Multi-centre radiomics for prediction of recurrence following radical radiotherapy for head and neck cancers: Consequences of feature selection, machine learning classifiers and batch-effect
Made with FlippingBook - Share PDF online