融合文字与标签的电子病历命名实体识别

准确的命名实体识别是结构化电子病历的基础, 对于电子病历规范化编写有着重要的作用, 而现今的分词工具对于专业的医疗术语无法做到完全正确的区分, 使得结构化电子病历难以实现. 针对医疗实体识别中出现的问题, 本文提出了一种在命名实体识别领域中改进的BiLSTM-CRF深度学习模型. 模型将文字和标签结合作为输入, 在多头注意力机制中使模型关注更多的有用信息, BiLSTM对输入进行特征提取, 得到每个文字在所有标签上的概率, CRF在训练过程中学习到数据集中的约束, 进行解码时可以提高结果的准确率. 实验使用人工标注的1000份电子病历作为数据集, 使用BIO标注方式. 从测试集的结果来看, 相对于传统的BiLSTM-CRF模型, 该模型在实体类别上的F1值提升了3%–11%, 验证了该模型在医疗命名实体识别中的有效性.     

Improving FEM crash simulation accuracy through local thickness estimation based on CAD data

In this paper, we present a method for estimating local thickness distribution in finite element models, applied to injection molded and cast engineering parts. This method features considerable improved performance compared to two previously proposed approaches, and has been validated against thickness measured by different human operators. We also demonstrate that the use of this method for assigning a distribution of local thickness in FEM crash simulations results in a much more accurate prediction of the real part performance, thus increasing the benefits of computer simulations in engineering design by enabling zero-prototyping and thus reducing product development costs. The simulation results have been compared to experimental tests, evidencing the advantage of the proposed method. Thus, the proposed approach to consider local thickness distribution in FEM crash simulations has high potential on the product development process of complex and highly demanding injection molded and cast parts and is currently being used by Ford Motor Company.     

基于Time-aware LSTM双向自动编码器的患者疾病分型

医学领域中, 患有相同疾病的患者之间也存在差异性, 看似简单的疾病也可能表现出不同程度的复杂性, 这给患者的识别、治疗和预后都带来巨大挑战. 本文使用以纵向非结构化时序存储的电子病历来解决患者异质性, 通过抓住就诊时间间隔不规律的特点增强对于隐藏信息的获取, 经过前向和后向的双向学习捕捉当前就诊记录与过去和未来信息的联系, 加深对于原序列特征提取的层次, 使模型做出更为精准的决策. 本文提出的BT-DST模型使用time-aware LSTM单元构造双向自动编码器学习患者强大的单一表示, 然后将其用于患者聚类, 通过统计分析得到患者针对当前疾病的亚型分型, 可针对不同群体采用不同类型的治疗干预, 为不同类患者提供针对其健康状况的精准医疗.