In this study, linear regression and neural network-based hybrid models are developed for modelling the daily ED visits. Month and week of the year, day of the week, and period of the day, are used as input variables of the linear regression model. Generated forecasts and the residuals are further processed through a multilayer perceptron model to improve the performance of forecasting. To obtain forecasts for daily number of patient visits, aggregation is used where the obtained periodical forecasts are summed up. By comparing the performances of models in generating periodical and daily forecasts, this chapter not only shows that hybrid model improves the forecasting performance significantly, but also aggregation fits well in practice.
TopIntroduction
Health services have become one of the largest industries globally and emergency departments (EDs) constitute the main component of these services. EDs are very important for health services as they provide prompt and essential care for patients (Sariyer, Taşar and Cepe, 2019). EDs are also determined as critical bottlenecks, since they are often the entry point for patients through a hospital for an unscheduled care (Holm and Dahl, 2011). Since visits to EDs occur without an appointment, unknown or random demand of these departments creates a great concern for the hospital management. Besides, increases in expectations from health services, average ages, and population rates continuously accelerate the number of visits to EDs. This growing demand for emergency services has created circumstances such that overcrowding in EDs is becoming a threat for public health (Taşar and Sariyer, 2018; Sariyer, Ataman and Kızıloğlu, 2018). Since providing prompt care, especially for life-critical situations, is vital for EDs having an unknown demand and an overcrowded environment, improving timeliness of emergency care should be main goal of service providers. Although, increasing resources should be a possible approach to improve timeliness of care, it is generally not recommended since it is not feasible and economic (Hoot and Aronsky, 2008). Thus, with the given amount of resources, effective demand management is indispensable in EDs to generate better plans for operations and consequent improved outcomes. Many researchers and policy makers have argued that for an organization to be able to generate efficient plans, anticipating the future is vital (i.e., Milner, 1997). Thus, generating real-like forecasts for number of ED visits can be used as a main parameter for demand management and efficient planning.
In literature, variety of techniques are proposed in forecasting demand in health services over time, for an exhaustive review see Gul and Celik (2018). Time series methods are most widely used in this context. For non-seasonal and stationary data, researchers prefer to apply simple techniques such as moving average and exponential smoothing in forecasting based on their practicability and superiority (i.e. Bergs et al., 2014; Ozudogru and Gorener, 2016; Sariyer, 2018a; Taylor, 2008). When the variability of the demand data increases, such as data is no-longer stationary, researchers require to apply more complex methods such as multivariate time series modelling (Jones et al., 2009), auto regressive integrative moving average-ARIMA (Hertzum, 2017, Sariyer, 2018b), or neural network models (Gul and Guneri, 2016; Xu, Wong and Chin, 2013). More recent approach used by researchers in forecasting demand of services is to use of hybrid models (Hadavandi et al., 2012; Yucesan, Gul and Celik, 2018).