Hospital readmission prediction based on long-term and short-term information fusion

Abstract

The hospital readmission prediction becomes a significant task for healthcare systems and patients. Many predictive models have been developed and make progress in this task. However, most of them roughly combine the patient’s long-term(e.g., the history of present illness) and short-term(e.g., the performed laboratory test results when the patients are discharged) information without considering the inner distinction between them. In this paper, we propose a new approach for hospital readmission prediction based on transformation from numerical features to natural language, which makes better fusion of these two kinds of information. Through a rule-based transformation, the original numerical features are transformed into corresponding descriptive short sentences based on medical knowledge. Meanwhile, with the help of public well pre-trained character embeddings, our model can incorporate the prior semantic knowledge into the data. Moreover, by using the long-term information as the query of short-term feature attention mechanism, our model can capture the effective information in the short-term features from a more global perspective, and better incorporates the long-term and short-term information. Extensive experiment results on a real dataset demonstrate the effectiveness and superiority of our proposed model compared with the baseline methods.