Predicting Incoming Hospital Patients Using Weather Data and Neural Networks

Sammanfattning: Personnel shortages in hospitals are a problem, and one way to alleviate that is to make better use of the existing personnel. This thesis examines whether Intensive Care patient arrivals can accurately be forecast using machine learning models, as well as examining the usefulness of calendar and climate data for the task. The two models used in this study were the Temporal FusionTransformer (TFT) and NeuralProphet (NP). The models were trained using hospital statistics from the city of Gävle, Sweden, as well as climate data from the same region. Results showed that calendar data such as weekday and the week of the year had highest importance on the forecast, but temperature, and specifically the lowest recorded temperature, were also contributing to the forecast. The TFT model forecasts were significantly better compared to the NP model, a mean absolute error of 0.5053 compared to 0.8322. This is in relation to data were the vast majority of target variable values ranged between 0 and 3. This shows some promise for the purpose of the task.