融合nmODE的术后肺部并发症预测模型  

Predicting postoperative pulmonary complications after lung surgery using nmODE

作　　者：熊立鹏 徐修远[1] 牛颢[1] 陈楠[2] 章毅[1] XIONG Lipeng;XU Xiuyuan;NIU Hao;CHEN Nan;ZHANG Yi(College of Computer Science,Sichuan University,Chengdu 610065,China;West China Hospital,Sichuan University,Chengdu 610065,China)

机构地区：[1]四川大学计算机学院,四川成都610065 [2]四川大学华西医院,四川成都610065

出　　处：《智能系统学报》2025年第1期198-205,共8页CAAI Transactions on Intelligent Systems

摘　　要：为了准确预测病人肺部手术后并发症的发生,提出了一种融合神经记忆常微分方程(neural memory ordinary differential equation,nmODE)的并发症预测模型。首先,利用极限梯度提升(extreme gradient boosting,XGBoost)树结构对数据进行编码,并提取其特征重要性。然后,使用长短时记忆神经网络对数据的相关特征依赖性进行分析,并提取处理后的特征。最后,利用nmODE的记忆和学习能力,对提取的特征进行深入分析,并得出最终的预测结果。通过实验评估,在肺部术后并发症数据集中,证明了提出模型的效果优于现有模型,同时可以为预测肺部手术后并发症的发生提供更准确的结果。In order to accurately predict the occurrence of postoperative complications in patients'lungs,a complication prediction model combining neural memory ordinary differential equation(nmODE)is proposed.The method of this model is as follows:firstly,an extreme gradient boosting(XGBoost)tree structure is used to encode the data and extract its feature importance.Then,a long short-term memory neural network is employed to analyze the dependency of the data's relevant features and extract the processed features.Finally,by utilizing the memory and learning capabilities of nmODE,the extracted features are deeply analyzed to obtain the final prediction results.Experimental evaluation has demonstrated the effectiveness of the proposed model in the dataset of postoperative complications in the lungs,showing superior performance compared with existing models.Furthermore,it can provide more accurate results for predicting the occurrence of postoperative complications in lung surgery.

关 键 词：疾病预测 异构表格数据 神经记忆常微分方程 极限梯度提升 长短时记忆神经网络 合成少数过采样技术 类别不平衡 病人预后 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]