4:03 PM Abstract No. 334 Reduction of unplanned late hours in inpatient procedure scheduling by forecasting with the Facebook Prophet algorithm

Inability to predict daily and monthly variation in interventional radiology (IR) inpatient needs is a major cause of unanticipated late hours and staff dissatisfaction. We tested the hypothesis that forecasting inpatient demand with the Facebook Prophet algorithm would reduce unplanned late inpatient hours compared to our standard scheduling method. A total of 12,483 inpatient cases performed between 2016 and 2019 were analyzed. For 2019, inpatient time was computed using two methods: (1) standard method (mean inpatient time from previous year) or (2) Facebook Prophet algorithm (a forecasting method based on an additive model where components of non-linear trend, seasonality, and holiday effects are complemented by a probabilistic model). Actual allocated inpatient time was performed using the Prophet forecast. We computed the difference between actual total inpatient hours performed and expected total inpatient hours using both methods from January to August 2019. One sample T-tests were used to compare deviations from expected values. Mean weekday inpatient hours were 22.9 and 23.3 in 2017 and 2018, respectively. Using the standard method, for 2019 the monthly planned inpatient time would be 23.5 hours. Using the Prophet algorithm, monthly planned inpatient time ranged from 24.1 (May) to 26.7 (August) hours, correctly anticipating both the upward trend, as well as the seasonality peak in August. Using the standard method, there was a significant deviation from the expected inpatient time in 2019 (P <0.001) in the months of March (-2.8 hours/d, P <0.001), April (-3.8 hours/d P <0.02), July (-3.5 hours/d, P <0.001), and August (-3.1 hours/d, P <0.02). Using the Prophet algorithm, there was no significant deviation from expected inpatient time in 2019 (P <0.052) or in the months of March (-1.1 hours/d, P <0.12), April (-1.2 hours/d, P <0.41), July (-0.68 hours/d, P <0.44), and August (-0.9 hours/d, P <0.93). Forecasting expected demand with the Facebook Prophet algorithm can significantly reduce unplanned inpatient hours, especially during months with high demand for inpatient procedures.