【作者】 张翼；

【导师】 黄晓林；

【作者基本信息】 南京大学 ， 电子与通信工程（专业学位）， 2020， 硕士

【摘要】 随着医疗信息技术的快速发展,重症监护室(Intensive Care Unit,ICU)集中越来越多的监护设备。目前临床监护设备的异常报警功能通常使用阈值检测法,采用多参数联合的超限报警以提升报警准确率。然而ICU误报警率依然很高,严重影响ICU治疗监护效果,这其中尤以心电监护仪误报警占主要。近几年,人们对综合使用ICU中多种生理监护数据以降低心电监护仪的误报警率的兴趣不断增加,随着生理信号分析算法的不断成熟与机器学习算法的进一步发展,二者的结合为ICU心律失常误报警率抑制问题提供了新的解决方案。ICU监护数据往往存在数据缺失、噪声干扰、信号种类复杂、数据集类别不平衡的问题,本次研究针对以上问题,着重于ICU多通道生理信号的预处理、心动节律信息的提取与融合、生理信号特征的分析、提取、筛选与组合,并采用随机森林与XGBoost两种机器学习算法建立五种严重心律失常:1.心动停搏(Asystole,ASY);2.心动过缓(Extreme Bradycardia,EBC);3.心动过速(Extreme Tachycardia,ETC);4.室性心动过速(Ventricular Tachycardia,VTA);5.室扑\室颤(Ventricular Fibrillation or Flutter,VFB)的误报警甄别模型,并探索各特征值组合、训练重采样方法与机器学习算法对不同类型心律失常误报警抑制的影响。本文的主要研究内容分为以下几个部分:一、数据集分析:对研究采用的ICU多通道生理信号数据集进行整体分析,对数据集中五种心律失常类别数量、正负样本分布情况、生理信号种类组合进行统计。然后根据信号数据特点选取机器学习算法并建立模型性能评估标准。二、预处理与心动节律信息提取:根据心律失常种类与监护仪设计标准提取关键时间段信号。而后针对数据集中心电信号(ECG)、动脉血压信号(ABP)与脉搏波信号(PPG)设计预处理方案,改良异常波形检测算法,并进行小波阈值工频降噪实验的探究。最后针对不同种类生理信号进行心动节律信息的提取,特别设计了基于生理信号指数的融合心率估计与ECG R峰结果校正算法。三、特征提取:综合考虑五种心律失常病理特征、多种生理信号特点,提取心率、R峰最大间期等时域特征;频率带宽、主频、均值频率等频域特征;间期周期性、峰值锐度等信号质量特征。并利用接收者操作特性曲线下面积(ROCAUC)进行单一特征的分类性能讨论。四、模型构建与结果分析:根据单一特征的分类性能指标AUC计算abs(AUC-0.5)区间,以进行特征组合。针对样本集正负类别不均衡问题,采用七种重采样方法,并通过对比实验,确定最适合各类样本的重采样方法,并将最优特征组合方案与重采样方法结合,应用于随机森林与XGBoost模型对比研究中。本研究在Physio Net Challenge 2015多通道生理信号数据集上实验了误报警甄别模型,最佳结果显示五种心律失常类型误报警抑制AUC均大于0.9,平均Score为81.9。VFB类型取得最高AUC值0.963,ETC类型取得最高Score值96.4。总计准确率为88.2%,平均AUC为0.941,总Score为81.7。在数据集较小的情况下,较Physio Net Challenge 2015第一名,Score提升0.31,敏感度提升3.0%。更多还原

【Abstract】 With rapid development of medical information technology,the intensive care unit(ICU)has been equiped with more and more physiological monitoring devices.At present,the multi-parameters threshold detection method is used to imporve accuracy of monitor’s alarm.However,the increasing false alarm rate leads to the decrease of monitoring effects.In recent years,increasing effort has been made to reduce the false alarm by comprehensively using multiple physiological monitoring data,especially combined with machine learing.ICU monitoring is always confronted by data missing,noise interference,complex signal types,and data category imbalance.Therefore,I will focus on the preprocessing of multi-channel biosignals,extraction and fusion of cardiac-rhythm information,extraction and combination of features.At the same time,Random Forest and XGBoost are applied to predict five severe arrhythmias: 1.Asystole(ASY);2.Extreme Bradycardia(EBC);3.Extreme Tachycardia(ETC);4.Ventricular Tachycardia(VTA);5.Ventricular Fibrillation or Flutter(VFB).I will investigate the effects of various features combinations,different resampling methods and machine learning(ML)algorithms on the false alarm reduction.The main research content is divided into the following parts:1.An overall analysis of the ICU multi-channel signal data was performed,including examining the sample size of five arrhythmia types,distribution of positive and negative categories,and multi-channel biosignal combinations.Based on these statistically analyses,the article selected appropriate ML methods and established the model evaluation standard.2.According to arrhythmia types and monitor standards,a key period extraction scheme was used.Then a multi-channel biosignal preprocessing method was applied to ECG,ABP and PPG signal.After that,the cardiac rhythm information was extracted from different biosignals.A heart rate estimation and R peaks correction algorithm was designed based on signal quality index.3.According to the five arrthythmias and three kinds of biosignals,I extracted time-domain features,such as heart rate,RR maximum intervals.Frequency-domian features,such as bandwidth,mean frequency,were also extracted.Some features are used to indicate signal quality,such as periodicity and sharpness measure.And the AUC of signal feature classification was analyzed and applied to the subsequent studies.4.ROC AUC was used to measure the classification performance.Then,a features combination scheme was designed,based on the result of abs(AUC-0.5).Seven resampling methods were adopted to resovle the problem of imbalanced categories,and the most suitable resampling method for each arrhythmia was determined by experiments.Finally,the features combinations and resampling methods were applied to the comparative study of random forest and XGBoost.This study test the effect of false alarm reduction model on Physio Net Challenge 2015 multi-channel signal dataset.The best results showed that the five types false alarm reduction AUC were all greater than 0.9,and the mean of Scores was 81.9.The VFB had the highest AUC: 0.963,and the ETC had the highest Score: 96.4.The total accuracy was 88.2%,average AUC was 0.941 and the total Score was 81.7.Even though the dataset was small,compared with the first place in Physio Net Challenge 2015,the Score increased by 0.31,and the Sensivity increased by 3.0%.更多还原