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

2026

Read Large Language Models for AI-Assisted Radiotherapy Scheduling: A Feasibility Study Under Realistic Operational Constraints (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-12896.txt.

Introduction

Zhang and six colleagues test whether a large language model can draft radiotherapy schedules that respect clinical and operational rules. In a simulated one-year setting with three treatment machines and 1,400 patients, the model produced schedules that passed rule-based checks. About 99% of treatment fractions stayed within the preferred weekly time window, and added objectives cut machine switching and Friday idle gaps sharply. The system also flagged impossible requests and suggested fixes.

Catalog record

YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE60885
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.12896

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Large Language Models for AI-Assisted Radiotherapy Scheduling: A Feasibility Study Under Realistic Operational Constraints Abstract Background: Radiotherapy (RT) patient scheduling is a complex operational problem involving multi-fraction treatment continuity, machine eligibility, first-fraction workflows, and competing patient- and clinic-centered objectives. Current scheduling processes often rely heavily on manual coordination and can be difficult to adapt in response to changing clinical demands. Purpose: This study investigates the feasibility of using a large language model (LLM) to generate candidate RT patient schedules satisfying predefined clinical and operational constraints. Methods: A simulated RT scheduling environment containing three medical linear accelerators (LINACs) was developed over a one-year period using synthetic patient arrivals and treatment characteristics modeled after clinical practice. A total of 1,400 new patients spanning 12 treatment categories with distinct scheduling requirements were generated. An LLM-based scheduling framework was implemented using structured natural language prompts encoding clinical rules, operational constraints, and scheduling objectives. Performance was evaluated across multiple scheduling scenarios, including weekly time-consistency baseline, LINAC continuity, gap-constrained temporal relaxation, and infeasible scheduling request handling. Generated schedules were independently validated using automated and manual verification procedures. Results: In the evaluated scenarios, the LLM-generated schedules satisfied the feasibility rules. Approximately 99% of evaluated fractions remained within a preferred 60-minute weekly treatment-time consistency window. Adding a LINAC-continuity objective reduced LINAC switching from 54.6% to 10.1%, and adding gap-constrained temporal relaxation reduced the Friday mean daily gap time from 169.5 to 89.2 minutes while maintaining approximately 99% of fractions within the preferred 60-minute weekly window. The framework also demonstrated robust handling of infeasible scheduling requests by explicitly identifying constraint violations and proposing interpretable corrective actions. Conclusion: This study demonstrates the potential feasibility of using LLMs to support RT patient scheduling in a realistic, constraint-rich simulated clinical environment. The results support further investigation of LLM-assisted scheduling as a flexible, human-in-the-loop decision-support approach for the patient scheduling task. Eric Zhang 1,2 , Wen Li 1 , Youfang Lai 1 , Annette Souranis 1 , Georgia Paparoidamis 1 , Michael Roumeliotis 1 , Xun Jia 1 1 Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University 2 Westlake High School, Westlake Village, California Keywords: radiotherapy, patient scheduling, large language model, clinical decision support I . Introduction Radiation therapy (RT) patient scheduling represents a complex resource allocation problem. Unlike general outpatient appointment scheduling, RT scheduling must coordinate multi-fraction treatment courses delivered over several weeks while satisfying numerous clinical and operational requirements 11 , including designated time slots for extended first-fraction treatments, minimum spacing requirements for multi-fraction regimens, treatment machine eligibility restrictions, and machine and staffing availability. These requirements directly affect RT resource utilization. Advances in RT practice further emphasize the need for efficient resource planning. For example, the increasing adoption of hypofractionated treatment regimens can substantially alter treatment capacity requirements 18 . Adaptive RT introduces additional scheduling complexity because treatment replanning, physician review, quality assurance must be coordinated within clinically relevant time windows 21 , 14 . In current clinical practice, scheduling remains largely manual and is performed by experienced administrators and therapists. This process is labor-intensive, cognitively demanding, and difficult to scale as patient volumes increase or operational conditions change. As a result, manual scheduling may limit efficient use of treatment resources when competing clinical and operational constraints must be balanced simultaneously. The complexity of the RT scheduling problem has motivated extensive research within operations research frameworks. A comprehensive review highlights the broad application of operations research methods for logistics optimization and resource planning in RT across strategic, tactical, and operational levels 26 . Early optimization-based approaches formulated RT scheduling as a mathematical programming problem, notably through priority-based non-block scheduling models designed to improve treatment allocation while accounting for patient urgency 2 . More recent studies have employed exact optimization models, such as integer programming, constraint programming, and column generation 4 , 5 . The successful translation of these theoretical operations research models into clinical practice has been demonstrated by implementing frameworks that optimize scheduling while strictly respecting clinical time window preferences 25 , 23 . Meanwhile, to address the inherent unpredictability of patient inflows, researchers have developed stochastic and simulation-based dynamic optimization strategies, showcasing the efficacy of adaptive decision-making under uncertain demand 13 , 6 , 24 . More recently, data-driven and machine learning approaches have gained traction in this domain. Predictive models were employed to estimate RT treatment durations, enabling more efficient workday divisions, better treatment scheduling management and device utilization 1 , 27 . An online dynamic patient scheduling framework was introduced to learn the relationships between patient features and actual wait times, dynamically adapting scheduling decisions to outperform traditional flat reservation policies 16 , 12 . Despite these advances, clinical translation can remain challenging because scheduling objectives, soft constraints, and exception-handling rules must be explicitly encoded and maintained. When local policies change or unexpected operational events occur, optimization models may require reformulation, parameter adjustment, or additional implementation effort. Recent advances in artificial intelligence (AI) have increasingly influenced medical physics and RT 20 , 19 , 15 . Within this landscape, large language models (LLMs) have emerged as versatile tools capable of interpreting complex natural language instructions, synthesizing structured data, and executing multi-step reasoning. LLM-based methods have been investigated for tasks such as treatment target delineation, patient-facing question answering, treatment planning, and structure-name standardization 17 , 28 , 8 , 10 , 7 . Beyond these RT applications, LLMs show promise in automating broader healthcare workflows, including administrative coordination, documentation, and digital task automation 22 , 29 . These capabilities are uniquely suited to the RT scheduling problem, a domain where patient data, institutional policies, and physician preferences must be integrated to yield feasible decisions. In contrast to conventional solvers that require formal mathematical encoding, LLMs may provide a complementary interface for translating heterogeneous natural-language policies, structured patient data, and operational preferences into candidate scheduling decisions. This inherent flexibility is particularly advantageous for RT scheduling, where operational logic is frequently procedural and nuanced rather than strictly algebraic. In this work, we investigate the use of LLMs for the direct generation of RT patient schedules within a framework of complex clinical and operational constraints. By integrating structured patient data, existing schedule states, and explicit constraint descriptions, the LLM is tasked with synthesizing comprehensive daily treatment schedules satisfying predefined feasibility rules while balancing resource utilization and operational preferences. To our knowledge, this study represents one of the first systematic evaluations of LLMs in RT scheduling under realistic simulated conditions. Our results indicate that, under structured prompting and simulated conditions, LLMs can generate schedules satisfying complex predefined feasibility rules while adapting to different operational objectives. II . Methods II . A . Clinical Environment, Patient Categories, and Scheduling Constraints The simulated scheduling environment was modeled after a three-LINAC workflow in our RT department. To define the clinical scheduling workflow, operational constraints, and representative treatment categories, we consulted with our radiation therapist team and summarized routine scheduling practices used in daily clinical operations. Treatments are delivered within a standard clinical operating window from 7:00 AM to 5:00 PM, corresponding to 600 available minutes per LINAC and 1,800 total machine-minutes per day. Under typical operating conditions, the system accommodates approximately 80–100 patients and 1,400–1,800 treatment minutes per day. To capture the

Read the complete text