AI-built surgery call scheduler passes constraint checks in simulation

A rule-based surgery residency call scheduling system developed with two agentic coding assistants met all three prespecified call-related Accreditation Council for Graduate Medical Education constraints in 99.2% of 12,000 service-month schedules generated in a computational validation study. The system did not use an AI model to assign shifts directly: Claude Code and Codex helped a general surgery resident write and debug conventional Python software, while a separate optimization engine and validator generated and checked the schedules. In 100 independent simulated academic years, 45 runs required no manual intervention in any of their 120 service-months, and no run required intervention in more than four months.

The proof-of-concept study modeled one general surgery residency program with a 24-hour in-house call structure. Its validation roster included 32 categorical general surgery residents, two preliminary general surgery residents, 20 off-service rotators from emergency medicine, orthopedic surgery and urology, and one colorectal surgery fellow. These participants were distributed among 10 schedule-generating call pools. Fairness and annual workload analyses covered 31 in-program surgery residents: 29 categorical residents and two preliminary residents. Three categorical residents who spend about half the year on a critical-care fellowship were excluded from per-resident workload comparisons because their intentionally reduced call footprint could distort the results.

Each full-year simulation ran from July 1 through June 30 and covered every month and call pool. The same total number of resident vacation weeks was redistributed in every run, with each resident receiving one randomly placed vacation week per semester, expanded into a Friday-anchored nine-day absence. The design allowed the researchers to test different combinations of absences without repeating one schedule 100 times. A unit of analysis was the service-month, meaning the schedule for one call pool during one calendar month, because that is the unit at which residency programs normally produce and publish monthly schedules. No human subjects or protected health information were involved.

The scheduling engine uses a two-phase, weekend-first procedure. It first samples 100 possible weekend layouts, assigning Friday, Saturday and Sunday according to each pool’s configuration, and filters them by compliance. Surviving candidates are ranked using a composite score based on the range of weekend assignments among residents, the range of weekend days worked, consecutive-weekend pairs and variation within postgraduate-year groups. The system then assigns weekdays chronologically. It favors residents with longer intervals since their previous call, followed by those with fewer shifts so far that month, and uses random tie-breaking only when other criteria are equal. A local exchange step tests pairwise weekday swaps and retains changes that improve the fairness score.

Every completed schedule is checked by a validation module separate from the generator. The tested call-related rules prohibited consecutive scheduled calendar days across pools, required at least two fully free weekends in every rolling 28-day window, limited in-house call to no more frequently than every third night when averaged over four weeks, and required 14 hours of post-call rest. Weekend triad assignments were treated as an exception to the consecutive-day rule because they do not represent in-house 24-hour call. The software records every candidate iteration, allowing a schedule to be reproduced from its seed and providing a trace of how each assignment was reached.

The system met all prespecified call-related rules in 11,906 of the 12,000 service-months, or 99.2%. The other 94 months were explicitly flagged as unresolved rather than released as non-compliant schedules. Two authors independently reviewed 120 randomly selected months without seeing the validator’s determinations, and their judgments agreed with it in all 120 cases. The unresolved months were concentrated in two of the 10 pools and were associated with complex vacation overlap, depleted eligibility or coordination between residents shared by different pools. The researchers did not prove that these cases were mathematically impossible; a larger search budget or a different solver might resolve them.

Fairness was measured as the difference between the most and least assigned residents within each pool and month. In 65.4% of service-months, that monthly shift range was no more than one, while 29.3% had complete shift equity, meaning every eligible resident received the same number of shifts. The mean shift range was 1.84 shifts. Weekend distribution was more uniform: 42.5% of service-months had identical weekend allocations for all residents, and 86.6% differed by no more than one weekend; the mean weekend range was 0.71. Across a full academic year, residents averaged 18.65 weekends and 78.23 total call days. Differences among postgraduate-year groups were interpreted as consequences of service-level variation rather than evidence of algorithmic inequity.

Across the 100 full-year runs, 45 required no manual coverage in any month, 26 contained one unresolved month, 22 contained two, four contained three and three contained four. The median was one unresolved service-month per year. Mean time to generate a complete academic schedule was 108.8 seconds, with a standard deviation of 59.1 seconds; runs involving more difficult absence patterns required additional candidate search and cross-pool flexibility. The 100-run validation consumed about three hours of unattended computation on a standard workstation. Development and validation together took approximately 110 hours of clinician time over ten months, with 265 code revisions across about 80 working sessions.

The findings provide proof-of-concept evidence rather than evidence of clinical benefit or real-world deployment. The software was tailored to one program’s call pools, weekend formats, postgraduate-year eligibility rules and 24-hour call structure, and its transferability to other specialties or systems such as night float has not been tested. It also did not evaluate the ACGME 80-hour weekly maximum or the requirement for a 24-hour period free of all clinical and educational responsibility, because those limits depend on rotations, case volume, clinics, conferences, documentation and other activity absent from a call calendar. There was no prospective deployment, and resident satisfaction, perceived fairness and actual time savings were not measured. The study also did not formally stress-test extreme simultaneous absences, and holidays received no preferential weighting in the assignment algorithm.

Despite those limitations, the study demonstrates a different role for generative AI from asking a chat model to invent a schedule. The models accelerated implementation of software whose clinical rules, fairness weights and rescue policies were specified and reviewed by the clinical authors. The resulting tool could regenerate affected months after vacation changes, evaluate one-to-one swaps and refuse exchanges that would violate constraints, although these features were not formally stress-tested under extreme conditions. The central result is therefore a software-development and validation approach: clinicians familiar with a scheduling problem used agentic coding tools to build an auditable optimizer that completes nearly all simulated months and identifies the few that still need human intervention.

Subject of Research: Computational validation of an AI-assisted surgery residency call scheduling system

Article Title: Generating ACGME-constraint-aware surgery residency call schedule software with the assistance of agentic coding large language models

Article References: Lies, I. H., Rippy, M., Bhattacharya, S. D., & Giles, W. H. (2026). Generating ACGME-constraint-aware surgery residency call schedule software with the assistance of agentic coding large language models. Global Surgical Education – Journal of the Association for Surgical Education, 5(1), Article 204. https://doi.org/10.1007/s44186-026-00601-8

Image Credits: AI Generated

DOI: 10.1007/s44186-026-00601-8

Keywords: surgery residency, call scheduling, ACGME constraints, agentic coding, large language models, schedule optimization, resident workload, schedule fairness, medical education, Generating, ACGME-constraint-aware, surgery