基于双向记忆传导的ICD自动编码方法

ICD Automatic Encoding Method Based on Bidirectional Memory Conduction

目的基于深度学习技术,探讨国际疾病分类(ICD)自动编码的方法。方法提取50095例来自麻省理工学院下属重症监护医学信息数据集信息齐全患者的诊断报告和诊断编码,实现一种基于双向记忆传导机制的注意力卷积神经网络,将记忆传导能力引入层级设计的注意力卷积神经网络中。从顺序因果约束和层级关联角度建模,高效捕获诊断文本常见的长序列依赖语境,同时利用层级间特征,提升编码特征重用性。比较本文网络与其他改进网络的性能。结果测试集的平均Macro F1值、Micro F1值、Macro ROC-AUC、Micro ROC-AUC和P@5较其他网络都有所改善,平均预测1名患者的诊断编码只需0.05s。结论经过改进网络设计,能较为准确地实现ICD的自动编码,编码时间相比人工编码大幅缩短,实际应用中可提高编码的工作效率,降低人工编码工作量。

Objective This article proposes a deep learning based method to explore the accuracy of automatic encoding for the International Classification of Diseases(ICD).Methods Extract diagnostic reports and diagnostic codes from 50095 patients with complete information from the Massachusetts Institute of Technology’s intensive care medical information datasets,and implement an attention convolution neural network based on bidirectional memory conduction mechanism.Introduce memory conduction ability into the hierarchical design of the attention convolution neural network.Modeling from the perspective of sequential causal constraints and hierarchical correlations,efficiently capturing common long sequence dependent contexts in diagnostic texts,while utilizing inter hierarchical features to enhance the re-usability of coding features.Compare the performance of this network with other improved networks.Results The average Macro F1 value,Micro F1 value,Macro ROC-AUC,Micro ROC-AUC,and P@5 of the test set improved compared to other networks,with an average prediction time of only 0.05s for a patient’s diagnostic code.Conclusion After improving the network design,the automatic encoding of ICD can be achieved more accurately,and the encoding time is significantly shortened compared to manual encoding.In practical applications,the efficiency of encoding can be improved,and the workload of manual encoding can be reduced.