Large Language Models for AI-Assisted Radiotherapy Scheduling: A Feasibility Study Under Realistic Operational Constraints

TL;DR

This study explores the feasibility of using large language models to generate feasible radiotherapy schedules under complex operational constraints, demonstrating promising results in a simulated environment.

Contribution

It introduces an LLM-based scheduling framework that encodes clinical rules and constraints, showing potential for flexible, human-in-the-loop RT scheduling support.

Findings

LLM schedules satisfied all predefined feasibility rules.

Approximately 99% of fractions were within the 60-minute treatment window.

Adding specific objectives reduced LINAC switching and treatment gaps.

Abstract

Radiotherapy (RT) patient scheduling is a complex operational problem. Current scheduling often relies on manual coordination and can be difficult to adapt to changing clinical demands. This study evaluated the feasibility of using a large language model (LLM) to generate candidate RT patient schedules satisfying predefined clinical and operational constraints. A simulated three-LINAC RT scheduling environment was developed over one year using synthetic patient arrivals and treatment characteristics modeled after clinical practice. A total of 1,400 new patients across 12 treatment categories were generated. An LLM-based scheduling framework used structured natural-language prompts encoding clinical rules, operational constraints, and scheduling objectives. Performance was evaluated across scenarios involving weekly time consistency, LINAC continuity, gap-constrained temporal relaxation, and infeasible request handling. Generated schedules were validated using deterministic rule-based checks and manual review. LLM-generated schedules satisfied predefined feasibility rules in the evaluated scenarios. Approximately 99% of evaluated fractions remained within the preferred 60-minute weekly treatment-time window. Adding a LINAC-continuity objective reduced LINAC switching from 54.6% to 10.1%. Adding gap-constrained temporal relaxation reduced Friday mean daily gap time from 169.5 to 89.2 minutes while maintaining approximately 99% of fractions within the 60-minute window. The framework also identified infeasible scheduling requests and proposed interpretable corrective actions. These results suggested that LLMs may support RT patient scheduling in constraint-rich simulated clinical environments, motivating further investigation of LLM-assisted scheduling as a flexible, human-in-the-loop decision-support approach for RT operations.