Related Experiment Video
Updated: Apr 18, 2026
Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
Published on: December 6, 2024
Automating Rule-Compliant and Equitable Call Schedules for Orthopedic Surgery Residents With Artificial Intelligence
Prushoth Vivekanantha1, Marc Daniel Bouchard1, Jeffrey Kay1
1Division of Orthopaedic Surgery, Department of Surgery, McMaster University, Hamilton, ON, Canada.
Background:
Call schedule generation is a time-intensive administrative task for residency programs. Traditional manual approaches often require hours of computation and can be inflexible. Large language models (LLMs) offer an efficient and adaptable alternative, therefore the purpose of this study was to assess if generative pretrained Transformer 5.2 (GPT-5.2, OpenAI) combined with a deterministic Python rule-checker, can automate complex, rule-compliant, and equitable call schedules for orthopedic surgery residents.
Methods:
Ten month-long residency blocks from a single institution were modeled, including both nonbackup and backup months in which junior residents required senior resident coverage. GPT-5.2 was accessed via the OpenAI application programming interface and prompted to follow 14 scheduling rules reflecting local institutional policy. For each block, 3 consecutive schedules were attempted, yielding 30 total runs. Performance was assessed by the proportion of successful, rule-compliant schedules generated and fairness metrics (Jain and Gini indices). Efficiency metrics included total wall-clock time, attempt duration, and estimated cost.
Results:
For nonbackup blocks, all 15 runs (100%) produced rule-compliant schedules with no terminations. Mean (SD) Jain and Gini indices were 0.948 (0.024) and 0.119 (0.031), respectively. Mean (SD) wall-clock time was 236 (148) s, with a mean (SD) cost per run of $0.15 ($0.03) United States Dollars (USD). For backup blocks, 13 of 15 (86.7%) runs produced successful, rule-compliant schedules, however, all blocks produced at least 2 valid schedules. Mean (SD) Jain and Gini indices were 0.936 (0.025) and 0.132 (0.031). Mean (SD) wall-clock time was 448 (400) s and the mean (SD) cost per run was $0.24 ($0.08) USD.
Conclusion:
GPT-5.2 can automate the generation of complex, rule-compliant and equitable call schedules for orthopedic surgery residents within minutes at a low computational cost of less than $1.00 USD.
