Smart scheduling of arthroplasty surgery with machine learning and optimisation improves operating room utilisation.

Abstract

Total hip and knee arthroplasty (THA and TKA, respectively) are effective procedures but are costly and resource intensive. As such, scheduling can have a significant impact on hospitals. The aim of this study was to evaluate the efficiency of a surgical schedule generated using machine learning (ML) and mathematical optimisation to current gold-standard scheduling practices. All primary and revision TKA and THA cases between April 2012 and February 2022 from a single institution were included (n=15 267). Data was split by year into training/validation and test sets. Procedure-specific models were trained for predicting operative time. Integer linear programming optimisation of operating room (OR) utilisation using these ML predictions was compared with historic scheduling (using surgeon-surgery-specific rolling average values). Weekly simulated schedules were generated and compared based on OR underutilisation, overtime and cases completed. The neural network models performed the best for all four procedures (median MSE: 594.6). This was a 7.1% improvement in 15 min buffer accuracy compared with rolling average times. The ML-predicted and optimised schedule reduced OR underutilisation (p<0.0001) and increased the number of cases (p<0.0001). OR underutilisation was reduced by 56.2% (13.3 minutes/day), while only increasing overtime by 17.2% (3.6 minutes/day), compared with the rolling mean. Overall, there was a 6.1% decrease (31 OR days) to complete the cases in the test set. ML-predicted operative times and optimisation has the potential to reduce idle OR time and improve patient throughput. Approaches to surgical scheduling leveraging data maximises utilisation of existing resources.