A Deep Neural Network Based Non-linear Ensemble Framework for Time Series Forecasting of Emergency Department Patients Visits

• Main Authors: CHEN, LI-AN, 陳立安
• Other Authors: CHEN, TZU-LI
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/6cbgrb

碩士 === 輔仁大學 === 資訊管理學系碩士班 === 107 === Due to the convenience of medical treatment and the provision of National Health Insurance in Taiwan, there are often problems in which emergency department are often and emergency overcrowding occurs. This research will use the historical emergency visits data to predict the daily patient visits through a time series model for doctors and nurses to make the most appropriate head count arrangement. From the above motivations, this study will develop three research purposes. 1. Use a time series data to make predictions and evaluate the effectiveness of each prediction model. 2. Use a heuristic algorithm in the model to see if it will effectively improve the model. 3. Use the combined forecasting method to confirm whether the prediction accuracy of the model can be improved. This research will combine the linear and nonlinear combination prediction methods with the following five models. Autoregressive Integrated Moving Average model(ARIMA), Gray Prediction, Recurrent Neural Networks(RNN), Long Short-Term Memory(LSTM), Gated Recurrent Unit(GRU). In addition to combine five models to predict the numbers of emergency patients, the research will also be adjusted in each model by using Differential Evolution(DE). Using Mean Square Error(MSE), Mean Absolute Percentage Error(MAPE) evaluation indicators to judge the quality of the results. It can be known from the experimental results that in the single model, the traditional time series model has a poorer prediction effect than the cyclic neural network model. The addition of heuristic algorithm and combined prediction method can improve the prediction accuracy, and the result is better than the single model.