ORAHS2024
Abstract Submission

159. Nested boosting for probabilistic spatio-temporal forecasting of EMS demand

Contributed abstract in session HC-3: Emergency Medical Services, stream Regular talks.

Thursday, 14:00-15:30
Room: Room S2

Authors (first author is the speaker)

1. Mostafa Rezaei
Information and Operations Management, ESCP Business School

Abstract

Reliable forecasts of EMS demand can assist healthcare managers in improving both tactical and strategic decision-making processes. Prediction distribution estimates can offer significantly more value than point estimates of the mean but are inherently more challenging to obtain. This challenge is compounded by the sparse nature of EMS demand, which if not effectively addressed, can result in significant variance in estimation. In this paper, we introduce a novel boosting approach to probabilistic spatio-temporal forecasting. In contrast to other boosting methods, our approach allows the final ensemble to be represented as a single decision-tree structure, thereby enhancing interpretability for end users without compromising performance. Additionally, our approach can incorporate covariates such as special day indicators and weather data. Furthermore, it can simultaneously forecast for multiple interconnected time series, such as emergency calls to different departments. By utilizing over four years of data on emergency calls in Montgomery County, PA, we demonstrate the superior performance of our proposed approach compared to other high-performing machine learning models in forecasting various quantiles of future call volumes across different regions of the city and hours of the week.

Keywords

Status: accepted

Back to the list of papers