Identification and cause analysis on unplanned reoperations by text classification approach

Unplanned reoperations (URs) not only increase the hospitalization period and healthcare cost, but also raise the death risk of patients. The analysis of URs is thus significant for their quality control and reduction. However, the massive text data generated in hospitals makes the identification of URs a tedious task with potential bias. Current research on UR is limited to data analysis and lack automated classification using deep learning and natural language processing. Here we propose the UR-Net framework. It implements the UR identification and cause analysis by processing the long texts of ward round documentation and applying few-shot learning on multi-class cause classification. Our framework consists of the URNet-XL with a batch fusion method based on XLNet model, and the URNet-GT for cause classification based on the pre-trained model combined with feature extraction modules of multi-head attention and a bi-directional Gated Recurrent Unit. High weighted F1 scores of 96.34% and 93.37% are obtained for the respective processes in comparison with the baseline methods. The Area Under receiver operating characteristic Curve (AUC) of 97.86% indicates an excellent UR classification on the unbalanced dataset. Our approach provides a new route of UR identification and analysis with the potential of reducing its occurrence.