RadOnc-Agent coordinates radiotherapy workflows with a large language model

Summary

RadOnc-Agent is a framework orchestrated by a large language model (LLM). The system routes clinical intent through 26 callable functions across four stages of radiotherapy. In synthetic and anonymized workflow tests, it achieved completion rates of approximately 96–99%. The authors position the results as a demonstration of technical feasibility rather than clinical validation.

Why it matters

Radiotherapy AI tools are typically isolated by tasks such as planning, adaptation, and follow-up. An orchestration layer that maintains longitudinal patient context could reduce handoff errors and the cognitive load on clinical staff. However, this study contributes a system integration rather than a new model. Its value depends on how much trust hospitals place in entrusting schema-constrained routing to an LLM orchestrator in a safety-critical oncology pipeline.

Highlights

• 26 callable functions across four clinical stages are invoked via a chat-based LLM supervisor using schema-constrained routing
• Ablation: Removing longitudinal state drops cross-stage completion from 96.5% to 84%; disabling schema/identity validation increases incompatible backend routing to 95.28%
• Evaluation covers 2,600 single-function calls, 200 synthetic cross-stage scenarios, and 120 workflow examples derived from 60 anonymized patient records

Critical take

The 'real' cohort of 60 patients and 120 examples is too small to draw any conclusions about clinical benefit. The authors also acknowledge that the study proves nothing about accuracy or prospective utility. More concerning is the 95.28% routing incompatibility rate observed when schema validation is removed. This reveals a fragile architecture. A single routing error could send a patient's plan to the wrong backend. This is a failure mode that a safety-critical oncology system can never tolerate. Yet the text does not mention any human-in-the-loop safety net.