Predicting the number of arrivals at the emergency department is a problem that is part of the broader field of time series forecasting, i.e. the prediction of future values. Several methods can be used in this field and for many years statistical analysis and inferences have been the method of choice. However, with the current growth in computing capabilities, the ease of access to powerful machines and the constant progress observed in the field of Machine Learning, neural networks and other artificial intelligence are becoming more and more appealing to solve this kind of problem and are currently the most powerful methods in terms of average accuracy of predictions.
For our prediction model of the number of arrivals at the emergency department, we have therefore chosen to rely on a Deep Learning model. More precisely, it is a multi-layer neural network composed of LSTM (Long-Short term memory) nodes and convolution operations. The combination of these two aspects allows the model to extract important information affecting the predictions while keeping the intrinsic dependencies of the sequential aspect of the available data.
Within the Machine Learning field, the computer tools currently available enable our model to have a high adaptability capacity. In other words, depending on the number of future values to be predicted, the expected granularity (day or hour), the different possible categories and the external factors having a potential influence on the predictions, our model will automatically adapt itself in order to return the best predictions and information according to the given objective.
However, since a neural network can often be too complex to interpret its behavior properly, we back up its results with a statistical analysis of the data in order to provide the user with a better understanding of the situation and of the model’s behavior.
As an example, we have implemented a prediction model trained on 2 years of data corresponding to the daily arrival numbers at the emergency department. The goal was to predict values 14 days in advance according to the age group of patients. The optimal model is obtained after the evaluation of over 300 neural networks, all with a different architecture and configuration. The predictions then returned by this model on test data reach an accuracy of 93.35%, i.e. about 187 out of 200 average daily arrivals.