ABSTRACT
A recurring theme during the pandemic was the shortage of hospital beds. Despite all efforts, the healthcare system still faces 25 % of resource strain felt during the first peak of coronavirus. Digitisation of Electronic Healthcare Records (EHRs) and the pandemic have brought about many successful applications of Recurrent Neural Networks (RNNs) to predict patients' current and future states. Despite their strong per-formance, it remains a challenge for users to delve into the black box which has heavily influenced researchers to utilise more interpretable techniques such as ID-Convolutional neural networks. Others focus on using more interpretable machine learning techniques but only achieve high performance on a select subset of patients. By collaborating with medical experts and artificial intelligence scientists, our study improves on the REverse Time AttentIoN EX model, a feature and visit level attention network, for increased interpretability and usability of RNNs in predicting COVID-19-related hospitalisations. We achieved 82.40 % area under the receiver operating characteristic curve and showcased effective use of the REverse Time AttentIoN EXTension model and EHRs in understanding how individual medical codes contribute to hospitalisation risk prediction. This study provides a guideline for researchers aiming to design interpretable temporal neural networks using the power of RNNs and data mining techniques. © 2022 IEEE.