Abstract
INTRODUCTION: Data-intensive machine learning is suitable for predicting the case durations of common surgical procedures. In contrast, Bayesian methods (e.g., applying the surgeon/scheduler's estimate) often perform well for uncommon procedures. Uncommon procedures dramatically affect case duration predictions, necessary for scheduling cases days to weeks before surgery. Procedure coding systems have changed over time, as has their disparate use, due to the large increase in ambulatory surgery. However, the current epidemiology of uncommon procedures is based on datasets from 25 years ago. We calculated contemporaneous incidence proportions for rare procedure combinations, those performed at facilities only once or twice per quarter. METHODS: The retrospective cohort study used de-identified, publicly available data from the 2010-2024 Florida ambulatory surgery databases, comprising distinct combinations of 20,014,189 cases distributed across 5,106,524 combinations of quarter, facility, and Current Procedural Terminology codes. There were also 11,643,813 cases in the 2009-2024 inpatient databases across 4,772,566 combinations of quarter, facility, and distinct combinations of International Classification of Diseases (ICD) procedure codes. RESULTS: The incidence proportions of procedures performed once or twice at each facility during the quarter performed, "doubletons," became progressively less common. In contrast, the change from ICD-9 to the more granular ICD-10-PCS in 2015 made singletons and doubletons more common for inpatient surgery. In 2024, approximately 66%, 78%, and 87% of procedures were observed just once or twice each quarter at the ambulatory surgery center, hospital outpatient department, and inpatient surgical suite where observed. These doubleton procedures accounted for approximately 18% of cases at ambulatory surgery centers, 36% at hospital outpatient departments, and 55% at inpatient surgical suites. Pooling hospital estimates, approximately 84% of procedures and approximately 44% of cases were among procedures performed just once or twice during the quarter. CONCLUSIONS: Although surgical procedure(s) are the most important predictors of operating room time, many procedures are performed rarely, resulting in little historical case duration data for case duration estimation. Freestanding ambulatory surgery centers have, for more than 10 years, remained different from hospitals in performing more common procedures. Hospitals should plan, when scheduling cases, that approximately 84% of distinct combinations of procedures and 44% of cases will have little to no procedure-specific historical data, and even less so by the combination of procedure and surgeon. Machine learning models for predicting case durations should therefore account for these uncommon procedures, for which Bayesian methods are well suited.