Automate Radiotherapy Patient Scheduling with Large Language Model
Abstract
Purpose
Radiation therapy patient scheduling is highly complex due to stringent clinical and operational constraints, such as treatment continuity over a multi-fraction course, time for first fraction, etc. Current scheduling practice is largely manual and inefficient, often resulting in suboptimal usage of treatment time. This work investigates the use of large language models (LLMs) to automate patient scheduling while preserving clinical feasibility.
Methods
Patient data were simulated based on statistical distributions derived from our clinical practice with three LINACs treating 80-100 patients daily. Patients were grouped into 12 categories sharing common scheduling requirements such as treatment time per fraction, first-fraction timing constraints, and machine eligibility. ChatGPT 5.2 was employed and prompts were designed to intake daily new patients to be scheduled given the treatment calendar filled with schedules for existing patients and generate updated treatment calendar satisfying all relevant constraints. Prompts were iteratively refined to resolve issues such as missed fractions and improper spacing of multi-daily treatments.
Results
LLM generated clinically feasible schedules that satisfied all clinical constraints. Using identical patient data over a 20-day horizon, five independent scheduling runs consistently produced feasible solutions, demonstrating robustness of the approach. The total unscheduled gap time across three LINACs ranged from 0 to 25 minutes, negligible relative to the total treatment time of 1600–1800 minutes. Additional evaluations over a 40-day horizon successfully addressed challenging scenarios, including fully booked first-treatment days, in which new patients were scheduled on the earliest feasible subsequent day, and linear accelerator failure scenarios, in which affected patients were rescheduled according to physician-provided priority rankings. Computation times were approximately 5 minutes.
Conclusion
This study demonstrates the feasibility of using LLM to replicate expert-level radiation therapy scheduling logic under complex clinical constraints. It shows promise for improving scheduling efficiency, adaptability, and resilience to disruptions.