Predicting COVID-19 and pneumonia complications from admission texts

Risk assessment for admitted patients is an important task especially so in times of pandemic when medical institutions and specialists are severely overburdened. Prioritising high-risk patients and deprioritising low-risk ones is of paramount importance to ensure the maximum effectiveness of limited medical resources. Admission reports contain a plethora of information. Combined with radiology reports and laboratory results received during the first 24 hours from admission this information is sufficient for experienced medical professionals to estimate the patient severity. Unfortunately at the peak of pandemic doctors not specialized in respiratory diseases are often forced to make such decisions which can lead to sub-optimal risk assessment and patient routing. In this paper we propose a novel method to predict the risk of death or the need for artificial lung ventilation (ALV) for patients hospitalized with pneumonia or COVID-19 based on their admission reports and possibly other information available in first 24 hours of hospital stay. We do it by applying a Longformer model with a combination of local and global attention, as opposed to more traditional deep learning methods (convolutional neural networks (CNN), recurrent neural networks (RNN) or Transformer models with global attention only). Our proposed approach has the following advantages in comparison with current deep learning - based risk assessment methods. First as it is shown in the paper the Longformer architecture outperforms the baselines including BERT and a combination of BERT and recurrent neural networks.