M. Jain1, M. Jafarnia2, N. Nayyar2,3, Y. Wang2, B. L. Gewertz1, R. Jain2,3 1Cedars-Sinai Medical Center,Department Of Surgery,Los Angeles, CA, USA 2University Of Southern California,Department Of Electrical Engineering,Los Angeles, CA, USA 3Vivace Systems,Los Angeles, CA, USA
Introduction:
Operating room time is one of the most expensive resources in the hospital. Effective surgery scheduling is crucial to reducing hospital costs. The greatest challenge in optimal surgery scheduling is the uncertainty in operative duration (ORD). Total ORD varies by procedure, nature of case (emergent vs. elective, primary vs. redo, etc.), surgeon, and a variety of other factors. The use of algorithms and machine learning methods can greatly improve the prediction of ORD and operating room utilization.
Methods:
To predict ORD by procedure and surgeon, we introduced 3 different models that estimate ORD, and then combined them to obtain a Hybrid Model of ORD. In Model 1, the estimate is based upon the historical data for ORD for a particular procedure performed by a particular surgeon. In Model 2, we combine the estimate obtained in Model 1 with each surgeon’s estimate of their own required ORD. Intuitively, Model 2 gives more weight to surgeons with more accurate personal ORD estimates. In Model 3, the estimate is based upon the historical data for ORD for a particular procedure performed by any surgeon. A Hybrid Model then chooses the best ORD predictor for each surgeon-procedure pair. We use machine learning techniques to train the model and derive ORDs for each surgeon and each procedure. We also evaluated the effects of patient age, gender, and BMI on ORD. To create the optimal surgery schedule, we introduced a Mixed Integer Linear Program (MILP) formulation that reduces overtime and idle time cost.
A data set of ORD from a tertiary care medical center between November 2013 and March 2016 was evaluated. Predicted ORD was deemed accurate if the actual ORD was less than 20% above or below the predicted ORD. The number of cases predicted accurately was calculated.
Results:
ORD prediction accuracy increased by 33% with use of the Hybrid Model (60% vs. 45%). Patient demographics such as age, gender, and BMI did not improve ORD prediction in the Hybrid Model. Overtime costs are reduced by nearly 66% with our ORD prediction model. Finally, this model demonstrates the potential to increase the overall case load by nearly 18% with no changes in other performance metrics.
Conclusion:
Algorithms and machine learning methods and can be used to improve surgery scheduling by using historical data to more accurately predict surgeon-specific and procedure-specific ORDs.