An Interpretable Disease Onset Predictive Model Using Crossover Attention Mechanism From Electronic Health Records
An Interpretable Disease Onset Predictive Model Using Crossover Attention Mechanism From Electronic Health Records
复制标题
使用电子健康记录中的交叉注意机制的可解释的疾病发作预测模型
DOI：
10.1109/access.2019.2928579
复制
发表时间：
2019-01-01
期刊：
影响因子：
3.9
通讯作者：
Kong, Lanju
中科院分区：
文献类型：
--
作者：
Guo, Wei;Ge, Wei;Kong, Lanju
分析患者的电子健康记录（EHR）可以帮助指导疾病预防和个性化治疗。因此，根据患者的EHR数据预测即将到来的就诊内的疾病发作信息（本文称为医疗代码）是一项重要任务。为了实现这一目标，必须解决EHR数据的实时性和高维性。此外，模型的预测结果必须是可解释的。现有的方法主要使用递归神经网络（RNNs）来建模EHR数据，并采用注意力机制来提供可解释性。然而，诊断信息和治疗信息通常被视为同一类信息，而忽略了两者之间的区别和联系。这导致对患者疾病发展的分析不清楚，预测结果不准确。为了解决这个问题，我们提出了一个交叉注意力模型（COAM）。该模型采用两个RNN分别处理诊断和治疗信息，并通过交叉注意机制利用两部分信息之间的相关性提高预测精度。它可以学习个人医疗诊断和治疗的有效表示，并提供可解释的预测结果。实验表明，COAM可以显着提高预测的准确性，并提供临床意义的解释。
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.