Artificial intelligence has made progress on individual radiotherapy tasks, yet these capabilities remain fragmented across clinical stages, software environments, and data modalities, preventing coverage of the full treatment-to-follow‑up pathway.
RadOnc-Agent formalizes radiotherapy into four clinical phases and exposes 26 callable functions through a conversational interface. A large‑language‑model (LLM) controller maps clinical intent to schema‑constrained calls, preserves patient and workflow context longitudinally, and routes requests to specialist services.
The evaluation comprises three experiments: 2,600 single‑function requests (7,800 repeat executions), 200 predefined synthetic cross‑stage scenarios (600 executions), and 120 workflow instances from 60 de‑identified patients (360 clean executions), covering decision‑to‑planning and planning‑to‑adaptation transitions.
RadOnc-Agent selected the intended function in 98.79% of single‑function executions, completed 96.50% of scripted cross‑stage workflows, and completed 96.67% of real‑patient workflow executions. Ablation studies showed that removing longitudinal state reduced cross‑stage completion to 84.00%, while disabling schema and identity validation increased mismatched backend dispatch from 0% to 95.28%.
These findings demonstrate the technical feasibility of an LLM‑orchestrated architecture for coordinating heterogeneous radiotherapy capabilities across longitudinal workflows, but they do not establish clinical correctness, utility, or prospective benefit.
Review