结合多模态融合方式的脉搏波房颤识别

针对心房颤动疾病诊断检测复杂,病理检查有创等问题,构建基于脉搏波与深度学习的心房颤动分类预测模型,实现对心房颤动疾病的准确预测。首先,通过脉搏波设备采集数据,与MIMIC-Ⅲ数据库数据共同构建PPG-AF数据集;其次,基于Pytorch深度学习框架构建用于房颤分类的ResNet-CBAM-1DCNN双通道卷积神经网络;最后,将数据集按照8:1:1的比例划分为训练集,验证集和测试集,将脉搏波和其对应的格拉姆角场图作为输入,通过对网络结构和超参数的优化,在测试集中分类的F1分数达到了97.30%,准确度达到98.12%。本研究基于脉搏波信号与双通道卷积神经网络模型,能够实现对心房颤动疾病的准确诊断,有望为临床医师制定最佳治疗决策提供重要依据。     

融合脑电与近红外脑地形图特征学习的多模式分类

脑地形图可以用来可以监测大脑的活动状态，为了准确提取被试大脑活动产生信号的空间特征以及有效提高分类准确率，结合脑地形图和卷积神经网络提出一种多模态脑地形图神经网络分类算法(MBTMNN),对运动想象和心算进行分类识别。对脑电和近红外信号进行预处理，提取脑电的能量特征和近红外中氧合血红蛋白浓度特征，结合各自电极位置统一所有样本的colormap后生成脑地形图，将二者同时输入到卷积神经网络并在特征层进行融合得到训练模型。利用2017年柏林脑电-近红外公开数据集进行六折交叉验证实验，数据集包含29名被试，各300个样本，在运动想象左/右、心算/静息、运动想象/心算/静息和运动想象左/右/心算/静息等4种分类场景中，分别达到了82.91%、94%、90.34%和78.18%的准确率，高于同数据集的近期研究和单模态方法。所提出方法能够有效融合脑电和近红外信号以提高分类精度。     

一种基于ECG的多层共轭对称Hadamard特征变换的房颤异常信号分类方法

目的:对心电图房颤异常信号进行检测和分析,利用多层共轭对称Hadamard特征变换模型,构建房颤异常信号分类系统。方法:采用多层共轭对称Hadamard特征变换的房颤识别方法,检测房颤异常信号分类特征。采用基于误差梯度反向传播Levenberg-Marquardt神经网络模型训练测试数据集。构建房颤异常信号分类器,并建立临床诊断分类模型。结果:该模型能有效提高特征分类效果,增加算法的收敛速度及计算精度,便于实时分析和诊断房颤异常疾病。结论:该模型能够捕获异常房颤信号的疑似波形,评估和分析信号特征,具有较高的系统鲁棒性。     

基于联合决策卷积神经网络的光学相干断层扫描图像自动分类

光学相干断层扫描(OCT)技术能实现视网膜的高分辨率三维层析成像,对视网膜疾病类型的诊断和发展阶段的分析具有至关重要的作用。临床基于 OCT 图像的视网膜疾病诊断主要依靠眼科医生对图像中病变结构的分析,这一人工分析过程不仅耗时而且易产生主观的误判。研究视网膜疾病的自动分析和诊断技术将极大减轻眼科医生的工作量,是实现高效诊疗的有效途径。针对视网膜OCT图像自动分类,构建一种联合决策的卷积神经网络分类模型。该模型利用卷积神经网络从原始输入OCT图像中自动地学习不同层级的特征,同时在网络多个卷积层上设计多个决策层,这些决策层能够根据网络中不同尺度的特征图分别对OCT图像分类,最后模型融合所有决策层的分类结果做出最终决策。在Duke数据集(3 231张OCT图像)上的实验结果表明,基于多层级特征联合决策的卷积神经网络分类模型对正常视网膜、视网膜年龄相关性黄斑变性和视网膜黄斑水肿的平均识别准确率达到94.5%,灵敏性达到90.5%,特异性达到95.8%。在HUCM数据集(4 322张OCT图像)上的实验结果表明,基于多层级特征联合决策的卷积神经网络分类模型的平均识别准确率达到89.6%,灵敏性达到88.8%,特异性达到90.8%。充分利用卷积神经网络中丰富的多层级特征,能够有效地对视网膜OCT图像实现准确的分类,为临床上视网膜疾病的辅助诊断提供技术支撑。     

基于CNN和频率切片小波变换的T波形态分类

心电实时监控是心血管疾病防治的重要手段。心电图中T波的变化是心肌缺血和心脏猝死等疾病的重要表征,T波形态自动识别是心电远程监控中一个重要问题。由于实时监护用心电的强噪声背景影响,传统的T波特征提取与分类算法遭遇瓶颈。提出一种结合切片频率小波变换和卷积神经网络的T波形态识别算法,包括:自动定位R波波峰位置与T波终点位置,从而确定一个包含有T波的片段;对该片段做频率切片小波变换,将生成的时频图像输入卷积神经网络,完成T波的形态分类。频率切片小波变换将信号转换到时频域上,呈现心电信号的时频能量分布特征;卷积神经网络的隐含层通过对时频图像进行3次卷积、激活与池化,完成时频图像的3次特征提取,这些特征具有平移、缩放不变性。使用欧盟ST-T数据库中的12 830个片段,采用3折交叉验证法来训练和测试卷积神经网络模型,最终使基于心拍的分类准确率达到97.34%,F1测度达到96.97%;基于样本实验的分类准确率为84.80%,F1测度为83.30%。模型在QT数据库测试的分类准确率为87.83%,F1测度为85.38%,泛化性能良好。对比其他T波分类算法(如决策树、支持向量机等),基于心拍实验的分类准确率提高1%～5%。研究结果证明,针对6类形态T波进行分类设计的算法不仅在分类准确率上有所提升,在鲁棒性和泛化性能方面也表现良好。另外,算法模型也适用于其他多种生理信号的分析,在医学图像分析领域也有一定的指导意义。     

基于深度残差卷积神经网络的心电信号心律不齐识别

心电图(ECG)信号在采集过程中容易受内部和外部噪声干扰,而且不同患者的ECG信号形态特征差异较大,即使同一患者在不同时间和环境下其ECG信号也会有差异,因此ECG信号特征检测与识别在心脏病远程实时监测与智能诊断中具有一定难度。基于此,本研究提出将小波自适应阈值去噪和深度残差卷积神经网络算法用于多种心律不齐的信号识别过程中。其中,使用小波自适应阈值技术完成ECG信号滤波,并设计了包含多个残差块(residual block)结构的20层卷积神经网络(CNN),即深度残差卷积神经网络(DR-CNN),对5大类心律不齐ECG信号进行了识别。然后,本文采用残差块局部神经网络结构单元构建DR-CNN,缓解了深层网络的收敛难、调优难等问题,克服了CNN随着网络层数增加而导致的退化问题;进一步引入批标准化(batch normalization)技术,保证了网络的平滑收敛。按照美国医疗器械促进协会(AAMI)的心搏分类标准,使用麻省理工学院和波士顿贝丝以色列医院(MIT-BIH)心律不齐数据库中94 091个ECG心搏信号(2个导联),完成了心律不齐多分类、室性异位搏动(Veb)和室上性异位搏动(Sveb)等分类识别实验。实验结果表明,本文所提出的方法在ECG信号多分类、Veb和Sveb识别中的准确率分别达到了99.034 9%、99.498 0%和99.334 7%。在相同的数据集和实验平台下,DR-CNN在分类准确率、特异性和灵敏度上均优于相同结构复杂度的CNN、深度多层感知机等传统算法。DR-CNN算法提高了心律不齐智能诊断的精度,该方法与可穿戴设备、物联网和无线通信技术相结合,可以将心脏病的预防、监测和诊断延伸到家庭、养老院等院外场景,从而提高心脏病患者的救治率,并且有效地节约医疗资源。     

基于金字塔卷积结构的深度残差网络心电信号分类方法研究

近年来,深度神经网络(DNNs)已广泛应用于心电图(ECG)信号分类领域,但是以往的模型从原始ECG数据中提取特征信息受限。因此,本文提出了一种基于金字塔型卷积层的深度残差网络(PC-DRN)算法,该算法中包含的金字塔型卷积(PC)层可以从原始ECG数据中同时提取多尺度特征,并采用深度残差网络训练ECG信号分类模型,可以实现对ECG信号的分类。本文使用2017心脏病学挑战赛(CinC2017)提供的公开数据集,验证本文提出方法对4类ECG数据的分类效果。本文选取精度和召回率之间的谐波均值F_1作为主要评价指标。实验结果表明,PC-DRN的平均序列级别F_1(SeqF_1)从0.857提升到了0.920,平均集合级别F_1(SetF_1)从0.876提升到了0.925。因此,本文提出的PC-DRN算法为ECG信号的特征提取和分类提供了一种新的思路,为心律失常的分类诊断提供了有效的手段。     

基于CNN算法的稳态体感诱发电位的特征识别

脑-机接口研究可为瘫痪病人的康复带来一种新的治疗方法。已有研究表明对手指或者正中神经施加一定频率的体感刺激,会引发相同频率且具有空间特异性的稳态体感诱发电位。为优化基于稳态体感诱发电位的脑-机接口的性能,通过快速傅里叶变换寻找12个健康被试的个人左手特定共振频率,采用事件相关谱扰动进行时频分析,检测其稳态体感诱发电位信号。基于共振频率对实验诱发的脑电信号进行1 Hz带通滤波,获得特定频带的数据,采用卷积神经网络(CNN)学习算法对其进行分类,并与采用共空间模式和支持向量机的特征提取及特征分类的方法(CSP+SVM)进行比较。所有被试的结果显示:基于共振频率滤波方法,采用CNN学习算法获得的离线分类准确率均高于85%,并且CNN学习算法的分类准确率显著性优于CSP+SVM的分类准确率(91.8%±5.9% vs 77.4%±8.5%,P<0.05)。因此,在基于稳态体感诱发电位的脑机接口的特征识别中,CNN学习算法相比传统使用的机器学习分类算法(如共空间模式+支持向量机)能够显著提升分类准确率,提高脑机接口的整体性能。     

基于极限梯度提升和深度神经网络共同决策的心音分类方法

心音是诊断心血管疾病常用的医学信号之一。本文对心音正常/异常的二分类问题进行了研究,提出了一种基于极限梯度提升(XGBoost)和深度神经网络共同决策的心音分类算法,实现了对特征的选择和模型准确率的进一步提升。首先,本文对预处理后的心音信号进行心音分割,在此基础上提取了5个大类的特征,前4类特征采用递归特征消除法进行特征选择,作为XGBoost分类器的输入,最后一类为梅尔频率倒谱系数(MFCC),作为长短时记忆网络(LSTM)的输入。考虑到数据集的不平衡性,本文在两种分类器中皆使用了加权改进的方法。最后采用异质集成决策方法得到预测结果。将本文所提心音分类算法应用于PhysioNet网站在2016年发起的PhysioNet心脏病学挑战赛(CINC)所用公开心音数据库,以测试灵敏度、特异性、修正后的准确率以及F得分,结果分别为93%、89.4%、91.2%、91.3%,通过与其他研究者应用机器学习、卷积神经网络(CNN)等方法的结果比较,在准确率和灵敏度上有明显提高,证明了本文方法能有效地提高心音信号分类的准确性,在部分心血管疾病的临床辅助诊断应用中有很大的潜力。     

基于卷积神经网络的先心病心音信号分类算法

心脏听诊是先天性心脏病(简称:先心病,CHD)初诊和筛查的主要手段。本文对先心病心音信号进行分析和分类识别研究,提出了一种基于卷积神经网络的先心病分类算法。本文算法基于临床采集的已确诊先心病心音信号,首先采用心音信号预处理算法提取并组织一维时间域上心音信号的梅尔系数转变成二维特征样本。其次,以1 000个特征样本用于训练和优化卷积神经网络,使用自适应矩估计(Adam)优化器,获得了准确率0.896、损失值0.25的训练结果。最后,用卷积神经网络对200个心音信号样本进行测试,实验结果表明准确率达0.895,灵敏度为0.910,特异度为0.880。同其它算法相比,本文算法在准确率和特异度上有明显提高,证实了本文方法有效地提高了心音信号分类的鲁棒性和准确性,有望应用于机器辅助听诊。     

Adaptive filtering, modelling and classification of knee joint vibroarthrographic signals for non-invasive diagnosis of articular cartilage pathology

Interpretation of vibrations or sound signals emitted from the patellofemoral joint during movement of the knee, also known as vibroarthrography (VAG), could lead to a safe, objective, and non-invasive clinical tool for early detection, localisation, and quantification of articular cartilage disorders. In this study with a reasonably large database of VAG signals of 90 human knee joints (51 normal and 39 abnormal), a new technique for adaptive segmentation based on the recursive least squares lattice (RLSL) algorithm was developed to segment the non-stationary VAG signals into locally-stationary components; the stationary components were then modelled autoregressively, using the Burg-Lattice method. Logistic classification of the primary VAG signals into normal and abnormal signals (with no restriction on the type of cartilage pathology) using only the AR coefficients as discriminant features provided an accuracy of 68.9% with the leave-one-out method. When the abnormal signals were restricted to chondromalacia patella only, the classification accuracy rate increased to 84.5%. The effects of muscle contraction interference (MCI) on VAG signals were analysed using signals from 53 subjects (32 normal and 21 abnormal), and it was found that adaptive filtering of the MCI from the primary VAG signals did not improve the classification accuracy rate. The results indicate that VAG is a potential diagnostic tool for screening for chondromalacia patella.     

Screening of knee-joint vibroarthrographic signals using statistical parameters and radial basis functions

Externally detected vibroarthrographic (VAG) signals bear diagnostic information related to the roughness, softening, breakdown, or the state of lubrication of the articular cartilage surfaces of the knee joint. Analysis of VAG signals could provide quantitative indices for noninvasive diagnosis of articular cartilage breakdown and staging of osteoarthritis. We propose the use of statistical parameters of VAG signals, including the form factor involving the variance of the signal and its derivatives, skewness, kurtosis, and entropy, to classify VAG signals as normal or abnormal. With a database of 89 VAG signals, screening efficiency of up to 0.82 was achieved, in terms of the area under the receiver operating characteristics curve, using a neural network classifier based on radial basis functions.     

融合CNN和BiLSTM的心律失常心拍分类模型

为更加准确地从动态心电中提取异常心拍,设计一种融合卷积神经网络(CNN)和多层双边长短时记忆网络(BiLSTM)的心律失常心拍分类模型。心电信号首先被分割成0.75 s和4 s两种不同尺度大小的心拍信号,然后利用11层CNN网络和3层BiLSTM网络分别对小/大尺度心拍信号进行特征提取与合并,并使用3层全连接网络对合并特征进行降维,最后利用softmax函数实现分类。针对MIT心律失常数据库异常心拍类型分布不均衡的问题,采用添加随机运动噪声和基线漂移噪声的样本扩展方法,降低模型的过拟合。采用基于患者的5折交叉检验进行模型验证。MIT心律失常数据库116 000个心拍的分类结果表明:所建立的模型针对4类心拍(正常、房性早搏、室性早搏、未分类)的识别准确率为90.42%,比单独使用CNN(76.45%)和BiLSTM(83.28%)的模型分别提高13.97%和7.14%。所提出的融合CNN和BiLSTM的心律失常心拍分类模型,相比单一基于CNN模型或者BiLSTM模型的机器学习算法,有更好的异常心拍分类准确率。     

基于肌电信号的膝关节跨越障碍角度预测方法

为解决人体跨越障碍物时膝关节角度输出的问题,针对性设计一种穿戴式信号获取实验台,对下肢运动姿态进行运动分析,将肌肉电信号及关节角度信号作为运动数据,对信号进行处理后利用BP神经网络预测跨越障碍时输出角度,提出一种利用BP神经网络算法,根据不同大腿抬起高度,分析膝关节运动主动肌与被动肌发力程度,预测输出人体跨越障碍时膝关节角度的方法,能够有效帮助假肢膝关节或康复机器人实现跨越障碍的复杂动作。     

Analysis of Vibroarthrographic Signals with Features Related to Signal Variability and Radial-Basis Functions

Knee-joint sounds or vibroarthrographic (VAG) signals contain diagnostic information related to the roughness, softening, breakdown, or the state of lubrication of the articular cartilage surfaces. Objective analysis of VAG signals provides features for pattern analysis, classification, and noninvasive diagnosis of knee-joint pathology of various types. We propose parameters related to signal variability for the analysis of VAG signals, including an adaptive turns count and the variance of the mean-squared value computed during extension, flexion, and a full swing cycle of the leg, for the purpose of classification as normal or abnormal, that is, screening. With a database of 89 VAG signals, screening efficiency of up to 0.8570 was achieved, in terms of the area under the receiver operating characteristics curve, using a neural network classifier based on radial-basis functions, with all of the six proposed features. Using techniques for feature selection, the turns counts for the flexion and extension parts of the VAG signals were chosen as the top two features, leading to an improved screening efficiency of 0.9174. The proposed methods could lead to objective criteria for improved selection of patients for clinical procedures and reduce healthcare costs.     

Computer-aided diagnosis of knee-joint disorders via vibroarthrographic signal analysis: a review

The knee is the lower-extremity joint that supports nearly the entire weight of the human body. It is susceptible to osteoarthritis and other knee-joint disorders caused by degeneration or loss of articular cartilage. The detection of a knee-joint abnormality at an early stage is important, because it helps increase therapeutic options that may slow down the degenerative process. Imaging-based arthrographic modalities can provide anatomical images of the joint cartilage surfaces, but fail to demonstrate the functional integrity of the cartilage. Knee-joint auscultation, by means of recording the vibroarthrographic (VAG) signal during bending motion of a knee, could be used to develop a noninvasive diagnostic tool. Computer-aided analysis of VAG signals could provide quantitative indices for screening of degenerative conditions of the cartilage surface and staging of osteoarthritis. In addition, the diagnosis of knee-joint pathology by means of VAG signal analysis may reduce the number of semi-invasive diagnostic arthroscopic examinations. This article reviews studies related to VAG signal analysis, first summarizing the pilot studies that demonstrated the diagnostic potential of knee-joint auscultation for the detection of degenerative diseases, and then describing the details of recent progress in analysis of VAG signals using temporal analysis, frequency-domain analysis, time-frequency analysis, and statistical modeling. The decision-making methods used in the related studies are summarized, followed by a comparison of the diagnostic performance achieved by different pattern classifiers. The final section is a perspective on the future and further development of VAG signal analysis.     

Representation of fluctuation features in pathological knee joint vibroarthrographic signals using kernel density modeling method

This article applies advanced signal processing and computational methods to study the subtle fluctuations in knee joint vibroarthrographic (VAG) signals. Two new features are extracted to characterize the fluctuations of VAG signals. The fractal scaling index parameter is computed using the detrended fluctuation analysis algorithm to describe the fluctuations associated with intrinsic correlations in the VAG signal. The averaged envelope amplitude feature measures the difference between the upper and lower envelopes averaged over an entire VAG signal. Statistical analysis with the Kolmogorov–Smirnov test indicates that both of the fractal scaling index (p = 0.0001) and averaged envelope amplitude (p = 0.0001) features are significantly different between the normal and pathological signal groups. The bivariate Gaussian kernels are utilized for modeling the densities of normal and pathological signals in the two-dimensional feature space. Based on the feature densities estimated, the Bayesian decision rule makes better signal classifications than the least-squares support vector machine, with the overall classification accuracy of 88% and the area of 0.957 under the receiver operating characteristic (ROC) curve. Such VAG signal classification results are better than those reported in the state-of-the-art literature. The fluctuation features of VAG signals developed in the present study can provide useful information on the pathological conditions of degenerative knee joints. Classification results demonstrate the effectiveness of the kernel feature density modeling method for computer-aided VAG signal analysis.     

Automated EEG signal analysis for identification of epilepsy seizures and brain tumour

Abstract

Electroencephalography (EEG) is a clinical test which records neuro-electrical activities generated by brain structures. EEG test results used to monitor brain diseases such as epilepsy seizure, brain tumours, toxic encephalopathies infections and cerebrovascular disorders. Due to the extreme variation in the EEG morphologies, manual analysis of the EEG signal is laborious, time consuming and requires skilled interpreters, who by the nature of the task are prone to subjective judegment and error. Further, manual analysis of the EEG results often fails to detect and uncover subtle features. This paper proposes an automated EEG analysis method by combining digital signal processing and neural network techniques, which will remove error and subjectivity associated with manual analysis and identifies the existence of epilepsy seizure and brain tumour diseases. The system uses multi-wavelet transform for feature extraction in which an input EEG signal is decomposed in a sub-signal. Irregularities and unpredictable fluctuations present in the decomposed signal are measured using approximate entropy. A feed-forward neural network is used to classify the EEG signal as a normal, epilepsy or brain tumour signal. The proposed technique is implemented and tested on data of 500 EEG signals for each disease. Results are promising, with classification accuracy of 98% for normal, 93% for epilepsy and 87% for brain tumour. Along with classification, the paper also highlights the EEG abnormalities associated with brain tumour and epilepsy seizure.     

Automatic de-noising of knee-joint vibration signals using adaptive time-frequency representations

A novel de-noising method for improving the signal-to-noise ratio of knee-joint vibration signals (also known as vibro-arthrographic (VAG) signals) is proposed. The de-noising methods considered are based on signal decomposition techniques, such as wavelets, wavelet packets and the matching pursuit (MP) method. Performance evaluation with synthetic signals simulated with the characteristics expected of VAG signals indicates good de-noising results with the MP method. Statistical pattern classification of non-stationary signal features extracted from time-frequency distributions of 37 (19 normal and 18 abnormal) MP method-de-noised VAG signals shows a sensitivity of 83.3%, a specificity of 84.2% and an overall accuracy of 83.8%.