Data Augmentation for Pathology Prioritisation: An Improved LSTM-Based Approach

Public hospitals receive large volumes of pathology results everyday. It is therefore challenging for doctors to comprehensively analyse all this data. Pathology data prioritisation would seem to provide at least a partial solution. It has been suggested that deep learning techniques can be used to construct pathology data prioritisation models. However, due to the resource required to obtain sufficient prioritisation training and test data, the usage of deep learning, which requires large labelled training data sets, was found not to be viable. The idea presented in this paper is to use a small seed set of labelled data and then to augment this data. The motivation here was that data augmentation had been previously employed successfully to address data scarcity problems. Four data augmentation methods are considered in this paper and used to train deep learning pathology data prioritisation models. Evaluation was conducted using Urea and Electrolytes pathology data. The results show a best recall and precision of 0.73 and 0.71 respectively.