An Interpretable Disease Onset Predictive Model Using Crossover Attention Mechanism From Electronic Health Records

Abstract

Analysis of patients' Electronic Health Records (EHRs) can help guide the prevention of diseases and personalization of treatment. Therefore, it is an important task to predict the disease onset information (referred to as medical codes in this paper) within the upcoming visit based on patients' EHR data. In order to achieve this objective, the real-time nature and high dimensionality of EHR data must be addressed. Moreover, the prediction results of the model must be interpretable. Existing methods mainly use Recurrent Neural Networks (RNNs) to model EHR data and adopt attention mechanism to provide interpretability. However, diagnosis and treatment information have usually been regarded as the same kind of information, the difference and relationship between the two parts being ignored. This has led to unclear analysis about the patient's disease development and inaccurate prediction results. To address this limitation, we propose a CrossOver Attention Model (COAM). This model adopts two RNNs to process diagnosis and treatment information, respectively, and then deploys a crossover attention mechanism to improve prediction accuracy by leveraging the correlation between the two parts of information. It can learn effective representations of personal medical diagnosis and treatment, and provide interpretable prediction results. Experiments demonstrate that COAM can significantly improve the accuracy of prediction and provide clinically meaningful explanations.

Ähnliche Arbeiten