Forecasting emergency department arrivals with neural networks

Abstract

Emergency departments often suffer from chronic overloading as well as seasonal spikes in number of arrivals. In this study three different Deep learning based models are used to try to predict the next days arrivals to the Pirkanmaa Hospital Districts emergency department. First one is a simple Recurrent Neural Network (RNN) which uses Long short-term memory (LSTM) cells for temporal relations. The latter two use a newly developed Temporal Fusion Transformer (TFT) architecture specifically made for multi-horizon time series forecasting. The dataset used to train the three different models contains a mix of variables thought to have effect on the end results predictions. The dataset is split into a training set and a validation set. The models learn from the data in the training set and their performance is measured using the validation set. The LSTM model is implemented using TensorFlow python library and the TFT model with PyTorch Forecasting. The results show that all three models perform better than the measured baseline. Although much faster to train, The LSTM model falls behind of the TFT architecture in terms of prediction accuracy. From the two TFT based models the hourly frequency model performs the best as it has access to more data. The TFT model achieves this performance advantage due to its more sophisticated network architecture and from the usage of temporal self-attention layers for learning long term dependencies. Due to the nature of its architecture the TFT model has better interpretability. From the Variable Selection Network we can deduce which variables contribute the most to the end results prediction and possibly even remove those that do not bring much value. The weights from the self-attention layer tells us which parts of the time series the model is focusing on. The use of machine learning, especially Neural Networks, is still quite a new phenomena in applications like in this study. The results from this study are promising and indicate that further research of the subject matter is warranted. Experimenting with different datasets and a wider range of ML methods could shed more light into the future of ED forecasting.