3690 - Development and External Validation of an Autonomous Multimodal Artificial Intelligence Agent for Adaptive Radiotherapy Decision-Making In Nasopharyngeal Carcinoma
Presenter(s)
S. Zhang1, G. Q. Zhou1, Y. Liu2, M. Hao1, X. Yu3, L. Jia4, Z. Wei5, H. Li2, and Y. Sun6; 1Department of Radiation Oncology, Sun Yat-sen University Cancer Center, State Key Laboratory of Oncology in South China, Guangdong Provincial Clinical Research Center for Cancer, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangzhou, China, 2Shenzhen United Imaging Research Institute of Innovative Medical Equipment, Shenzhen, China, 3Sun Yat-sen Memorial Hospital, Guangzhou, China, 4Shenzhen United Imaging Healthcare Co., Ltd., Shenzhen, Guangdong, China, 5Cooperation Innovation Department, Shanghai United Imaging Healthcare Co., Ltd, Shanghai, China, 6Sun Yat-sen University Cancer Center; State Key Laboratory of Oncology in South China; Collaborative Innovation Center for Cancer Medicine; Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangzhou, China
Purpose/Objective(s):
We hypothesized that an autonomous multimodal large language model (LLM)-based agent can accurately identify optimal adaptive radiotherapy (ART) triggering timing in nasopharyngeal carcinoma (NPC) by integrating anatomical, dosimetric, and biological information. We aimed to develop and externally validate this agent in multi-institutional cohorts.Materials/Methods:
Eighty-two NPC patients (2,706 fractions) treated between 2022–2025 were included as the internal cohort. Two independent external cohorts (22 patients, 726 fractions) were used for validation. We developed a consensus-driven multimodal ART decision agent using a Qwen3-235B LLM without fine-tuning. The agent architecture features a simulated expert committee, comprising distinct virtual specialists, including an Anatomy Expert, Dosimetry Expert, Radiobiology Expert (interpreting NTCP and EBV DNA kinetics), and a Knowledge Librarian (leveraging a real-time updated RAG knowledge base). An automated pipeline integrated automated contouring, IGRT, and treatment planning to enable fully automated computation for each treatment fraction, with a processing time of approximately 8 minutes per fraction. At each fraction, structured ART trigger recommendations were generated. Modality weights were assigned based on clinical expertise in a ratio of 3:3:2:1:1 for anatomical, dosimetric, NTCP, EBV DNA, and other factors, respectively. Gold-standard trigger timing was determined by consensus of two senior radiation oncologists. Temporal agreement was defined as predictions within a predefined ±3-fraction window. Performance was evaluated using accuracy, sensitivity, specificity, and Cohen’s kappa. Interpretability was assessed by three experts using five-domain (completeness, relevance, correctness, clarity and actionability) Likert scoring with intraclass correlation coefficient (ICC).Results:
The agent achieved 76.3% accuracy (sensitivity 72.5%, specificity 79.9%) with moderate agreement (? = 0.53). External validation demonstrated stable performance (accuracy 71.6%, ? = 0.51). Multimodal integration improved discrimination of clinically indicated ART compared with anatomy-only evaluation (?accuracy = 13.2%). Mean predicted trigger fraction was 11.7 versus 13.9 by expert consensus, reflecting a modest bias toward earlier intervention. Temporal agreement within ±3 fractions was achieved in 83% of patients; discordance was predominantly early (16%) with rare delayed triggers (1%). Interpretability was high (mean Likert 4.58/5, ICC 0.87).Conclusion:
An autonomous multimodal LLM-based ART decision agent demonstrated accurate trigger timing prediction and robust external validity in NPC. Multimodal integration enhanced decision consistency and interpretability, supporting scalable precision-guided adaptive radiotherapy workflows.