Lattice LSTM神经网络法中文医学文本命名实体识别模型研究

目的探索利用点阵(lattice)长短期记忆(long short term mermory network,LSTM)神经网络构建命名实体识别(named entity recognition,NER)模型解决中文医学文本的信息提取问题。方法利用Lattice LSTM来表征句子中的词汇词(lexiconword),从而将潜在词信息整合到基于字符的长短期记忆网络—条件随机场(long short term memory-conditional random?eld,LSTM-CRF)模型中。进一步使用一个大型自动获取的词典来匹配句子,进而构建基于词的Lattice。利用Lattice LSTM结构自动控制从句子开头至结尾的信息流。结果门控单元可用于将来自不同路径的信息动态传送到每个字符。在NER数据基础上进行训练后,LatticeLSTM能够学会从语境中自动找到更有用的词汇,以取得更好的NER性能。结论与基于字符和词的NER方法相比,本文所提出的模型优势在于利用显性词汇信息而不是字符序列进行标注,同时较少出现分词误差。

Exploring the link between brain topological resilience and cognitive performance in the context of aging and vascular risk factors:A cross-ethnicity population-based study

Brain aging is typically associated with a significant decline in cognitive performance.Vascular risk factors(VRF)and subsequent atherosclerosis(AS)play a major role in this process.Brain resilience reflects the brain’s ability to withstand external perturbations,but the relationship of brain resilience with cognition during the aging process remains unclear.Here,we investigated how brain topological resilience(BTR)is associated with cognitive performance in the face of aging and vascular risk factors.We used data from two cross-ethnicity community cohorts,PolyvasculaR Evaluation for Cognitive Impairment and Vascular Events(PRECISE,n=2220)and Sydney Memory and Ageing Study(MAS,n=246).We conducted an attack simulation on brain structural networks based on k-shell decomposition and node degree centrality.BTR was defined based on changes in the size of the largest subgroup of the network during the simulation process.Subsequently,we explored the negative correlations of BTR with age,VRF,and AS,and its positive correlation with cognitive performance.Furthermore,using structural equation modeling(SEM),we constructed path models to analyze the directional dependencies among these variables,demonstrating that aging,AS,and VRF affect cognition by disrupting BTR.Our results also indicated the specificity of this metric,independent of brain volume.Overall,these findings underscore the supportive role of BTR on cognition during aging and highlight its potential application as an imaging marker for objective assessment of brain cognitive performance.

神经系统疾病智慧分诊管理模型初探

目的探索运用人工神经网络及贝叶斯决策理论建立神经系统疾病门诊智慧分诊决策树管理模型。方法以贝叶斯决策理论为理论基础,以人工神经网络技术完成神经系统疾病快速专科/亚专科机器学习;针对神经系统疾病专科或亚专科分诊数据,以循环神经网络及贝叶斯算法完成神经系统疾病症状与诊断的概率分布及收敛,建立神经系统疾病决策树管理模型并完成理论论证。结果完成了神经系统疾病智慧分诊的管理理论及模型构建,根据迁移学习特性基本实现神经系统疾病的快速学习和精确分诊。结论该管理模型的研究,能为后续应用提供理论借鉴意义,并在一定程度上缓解目前患者退换号率较高的问题。