Forecasting Teleconsultation Demand with an Ensemble Attention-Based Bidirectional Long Short-Term Memory Model

Abstract

Accurate demand forecast can help improve teleconsultation efficiency. But teleconsultation demand forecast has not been reported in existing literature. For this purpose, the study proposes a novel model based on deep learning algorithm for daily teleconsultation demand forecast to fill in the research gap. Because of the significant effect of holidays on teleconsultation demand, holiday-related variables, and specific prediction technologies were selected to treat it. The technologies attention mechanism and bidirectional long short-term memory (BILSTM) were used to construct a novel forecasting methodology, i.e., ensemble attention-based BILSTM (EA-BILSTM), for the accurate forecasts. Based on actual teleconsultation data, the effectiveness of variable selection is verified by importing different inputs into models, and the superiority of EA-BILSTM is verified by comparison with nine benchmark models. Empirical results show that importing selected variables can lead to better forecasts and EA-BILSTM model can get lowest forecasting errors on two sub-datasets. This indicates that the proposed forecasting model is a high potential approach for teleconsultation demand prediction in the influence of sparse trait, like holiday effects.

Forecasting Teleconsultation Demand Using an Ensemble CNN Attention-Based BILSTM Model with Additional Variables

“…The missing data is treated by the feature-driven method in the previous study [3]. To make the forecasting results more reliable, two datasets are used in the experiments, namely the 528-day dataset (1 January 2018 to 12 June 2019) and the 699-day dataset (1 January 2018 to 30 December 2019).…”

“…To test the effectiveness of additional variables and the superiority of the proposed model, the ensemble attention-based BILSTM (EA-BILSTM) model, proposed in the previous study [3], is built and used as the benchmark. The EA-BILSTM has been proved to outperform nine methods, including the traditional econometric model ARIMA, four machine learning models, K Nearest Neighbor (KNN), Support Vector Regression (SVR), For EA-BILSTM, two inputs are matrix I 1 ∈ R 7 × 8 and matrix I 2 ∈ R 9 × 7 .…”

“…In these two inputs, the time lag of history data is 7. The sizes of input metrics are determined by the autocorrelation and partial correlation analyses and the selection of holiday-related variables [3]. For ECA-BILSTM, the parameters of BILSTMs, Dense, and Attention modules are the same as these of EA-BILSTM.…”

“…Accurate teleconsultation demand forecasts are important for the efficient utilization of limited teleconsultation resources. To build an accurate forecasting model for daily teleconsultation demand, an ensemble deep learning method was proposed in the previous study [3]. Before the previous study, there wasn't related literature reporting daily teleconsultation demand forecast.…”

“…In the previous study of teleconsultation demand forecasting, the proposed forecasting methodology is the ensemble attention-based bidirectional long short-term memory model (EA-BILSTM) [3]. This model was proposed to deal with the sparseness and variability of holiday influence.…”