Predicting Daily Surgical Volumes Using Probabilistic Estimates of Providers' Future Availability

Abstract

Probability-based models are developed using information from a variety of datasets to predict daily surgical volumes weeks in advance. The quest was motivated by the need to make real-time adjustments to staff capacity and reallocation of the operating room block time based on predicted future demand. We test the notion that more data always leads to better predictions. Four probabilistic prediction models are presented, each parameterized based on real data and information from different sources. We hypothesize that the accuracy of the prediction improves by incorporating additional information. Models are tested for a surgical service at a large hospital using data of 20 months (January 19, 2015-August 31, 2016). We find that incorporating additional information may not improve prediction accuracy if that information is prone to data errors. However, deploying analytical data treatment to ameliorate these errors leads to better predictions. We also compare the predictive ability of the probability-based models to neural network-based models and find that the neural network models do not outperform simpler models. Managers should critically review the accuracy of the data used in decision-making. While a greater amount of inherently error-free information is the best, analytics can enhance the utility of error-prone data.