The degree of accuracy in surgery duration estimation directly impacts on the quality of planned surgical lists. Model selection for the prediction of surgery duration requires technical expertise and significant time and effort. The result is often a collection of viable models, the performance of which varies across different strata of the surgical population. This paper proposes a prediction framework to be used after a comprehensive model selection process has been completed for surgery duration prediction. The framework produces a partition of the surgical cases and a "hybrid model" that allocates different predictors from the collection of viable models to different parts of the surgical population. The intention is a flexible prediction process that can reassign models and adapt as surgical processes change. The framework is tested via a simulation study, and its utility is demonstrated by predicting surgery durations for Ear, Nose and Throat surgeries in a New Zealand hospital. The results indicate that the hybrid model is effective, performing better than standard model selection in two of the three simulation studies, and marginally worse when the selected model was the true underlying process.
Keywords: Cross-validation; Hybrid model; Linear regression; Prediction; Simulation.