CCU数据的多模态融合在动态死亡风险预测中的应用  

Application of multimodal fusion of CCU data in dynamic mortality risk prediction

作　　者：李然 邱皖 娄岩[2] LI Ran;QIU Wan;LOU Yan(Information Engineering College,Dalian Ocean University,Dalian 116023,Liaoning Province,China;School of Intelligent Medicine,China Medical University,Shenyang 110122,China)

机构地区：[1]大连海洋大学信息工程学院,辽宁大连116023 [2]中国医科大学智能医学学院,沈阳110122

出　　处：《天津工业大学学报》2022年第6期76-82,共7页Journal of Tiangong University

基　　金：河南省科技厅科技攻关项目(212102210379);中国医药教育协会重大科学攻关问题和医药技术难题课题(2020KTE010);教育部行指委职业教育改革创新课题(HBKC217010)。

摘　　要：针对冠心病监护室(CCU)患者的死亡风险预测集中在重症监护室(ICU)的患者,并且CCU患者的大量多模态数据不能被充分地利用的情况,提出一种动态预测CCU患者死亡风险的多模态融合方法。首先分析了CCU多模态数据的特点,从结构和时间2个维度提出了一种分类模式,使用卷积神经网络(CNN)和文本数据聚类的方法处理心脏超声报告(ECHO),使用信号处理方法处理高频波形数据心电图(ECG),将这2种方法提取的特征与其他结构化数据相结合,通过加入时间维度的决策树模型进行预测,决策树模型根据时间戳区分不同时期生成的数据。在MIMIC-Ⅲ数据集上的实验表明:与基准的机器学习方法支持向量机(SVM)相比,本研究在AUC-ROC/AUC-PR上分别获得11.36%和25.31%的改进,与基于线性分类的机器学习方法相比,在AUC上获得11.42%的改进。For the prediction of mortality risk in the coronary care unit(CCU) focused on patients in the intensive care unit(ICU), a large amount of multimodal data of patients in the CCU cannot be fully utilized, and a multimodal fusion method for dynamic prediction of mortality risk of patients in the CCU is proposed. The characteristics of CCU multimodal data are first analyzed. A classification model is proposed in both structural and temporal di-mensions, using CNN and text data clustering methods for cardiac ultrasound reports(ECHO) and signal processing methods for high-frequency wave form data electrocardiograms(ECG), and combined the features extracted by these two methods with other structured data for prediction by adding a decision tree model with a temporal dimension, which distinguishes data that are generated in different periods according to the time stamps. Experiments on the MIMIC-Ⅲ dataset show that this study obtains 11.36% and 25.31% improvement on AUC-ROC/AUC-PR compared to the benchmark machine learning method of support vector machine(SVM),and 11.42% improvement on AUC compared to the linear classification-based machine learning method.

关 键 词：多模态融合 动态死亡率预测 冠心病监护室(CCU) 死亡风险预测 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]