Artificial intelligence has made significant strides in individual radiotherapy tasks, yet these capabilities often remain isolated across disparate clinical stages and software environments. A new research paper introduces RadOnc-Agent, an agentic AI framework designed to bridge these gaps by formalizing the radiotherapy pathway into four distinct clinical phases.
The Orchestration Framework
At the core of RadOnc-Agent is a large-language-model (LLM) controller that maps clinical intent to schema-constrained calls. The system provides 26 callable functions through a conversational interface, maintaining patient and workflow context throughout the longitudinal process. By routing requests to specialist services, the framework attempts to unify the workflow from initial treatment decision-making through to follow-up care.
Performance and Validation
The system's execution was evaluated through a rigorous multi-tier testing process:
- In 2,600 single-function requests, RadOnc-Agent selected the intended function with 98.79% accuracy.
- In 200 synthetic cross-stage scenarios, it achieved a 96.50% completion rate.
- Using 120 workflow instances derived from 60 de-identified real-patient records, the system successfully completed 96.67% of executions.
Comparative ablation studies highlighted the necessity of specific architectural features. Removing the longitudinal state tracking caused cross-stage completion rates to drop from 96.50% to 84.00%. Furthermore, disabling schema and identity validation resulted in mismatched backend dispatching in 95.28% of test cases, illustrating the importance of rigorous validation protocols.
Technical Feasibility vs. Clinical Utility
The researchers emphasize that while these findings establish the technical feasibility of using an LLM-orchestrated architecture to coordinate heterogeneous radiotherapy data, they do not yet prove clinical efficacy. The study does not establish the clinical correctness or prospective benefits of the framework in a live medical environment.
