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ART2神经网络的动态ECG数据压缩算法

采用ＡＲＴ（ＡｄａｐｔｉｖｅＲｅｓｏｎａｎｃｅＴｈｅｏｒｙ）神经网络中的适应于模拟信号输入的ＡＲＴ２网络结构，提出了一种动态ＥＣＧ数据压缩算法。借助于ＡＲＴ２优良的模式识别特性，该算法能实现高压缩比和高压缩精度的动态ＥＣＧ数据压缩。ＡＲＴ２网络能迅速识别已学习样本及事先未知的新样本，因此该算法不仅能快速适应各种未知ＥＣＧ信号的动态变化，而且提高了算法的运行速度和数据压缩比。     

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。     

ART神经无网络的警戒线参数的优化设计

具有广泛应用潜力的ＡＲＴ网络的主要缺陷之一，是其警戒线参数的取值大小完全依赖使用者的经验，迄今无有理论及方法上的指导；本文分别通过蒙特卡洛抽样来获取大量模式样本、并应用模糊聚类的优化分割技术进行最佳分类；然后，在最佳分类的基础上，将模式样本序贯输入ＡＲＴ网络，同时调整警戒线参数，直至建立起准确的分类器；最后，将不同分类数下对应的譬戒线参数拟合成经验公式，从而为给定分类数下较准确获取警戒线参数解决了重要的技术困难。     

心音信号的判别分析

本文用自回归模型提取心音信号的特征,用判别分析的方法在微型计算机上对278个心动周期作了二类(正常与不正常)自动分类,取得了95％的分类率。文中重点讨论了心音特征选择的必要性及各种选择的方法和取得的相应结果。实现了心音信号的实时两类判别。     

基于交叉对比神经网络的心音分类

目的:通过交叉对比神经网络（CCNN）实现心音信号的自动分类,从而对心血管疾病进行早期诊断。方法:实验基于PhysioNet/Cinc 2016心音数据库。训练集和测试集数据来自互斥的健康受试者/病理患者,并以4:1的比例进行划分,输入CCNN。CCNN利用深度卷积神经网络进行特征提取,结合基于信息的相似度度量理论（IBS）,对特征向量间的相似性进行度量并分类。结果:实验结果得出灵敏度为0.834 6,特异性为0.962 3,最终大赛综合得分为0.898 5。结论:CCNN使用交叉对比的输入模式扩充数据量,引入信号间的对比信息,同时在神经网络的训练过程中应用统计学思想,使网络具备良好的泛化性,更加适应医学数据量较少的场景,在心音分类中取得较好的结果。     

基于时频组合特征与自适应模糊神经网络的心音分类算法

特征提取方法和分类器的选择是心音分类中的两个重要环节。为了充分捕捉心音信号中的病理性特征,研究中引入了一种结合梅尔频率倒谱系数（MFCC）和功率谱密度（PSD）的特征提取方法。与目前常规分类器不同,研究中选择了自适应模糊神经网络（ANFIS）为分类器。在实验设计方面,选取了不同时期、不同频率范围的PSD进行对比,选出分类效果最佳的特征,并采用均值PSD、标准差PSD、方差PSD和中位PSD四种不同的功率谱统计特性进行对比。通过实验比较,心音收缩期100～300 Hz的中位PSD和MFCC组合特征有最好的效果,在准确率、精确率、灵敏度、特异度和F1得分上分别达到96.50%、99.27%、93.35%、99.60%和96.35%。结果显示本研究所提算法对先心病辅助诊断具有较大帮助。     

基于机器学习的心音识别分类研究

心音信号可反映心脏的病理信息,是诊断心脏健康的重要依据之一。本文首先从心音信号提取时频域、梅尔倒谱系数等145个特征作为机器学习的输入数据集,然后在随机森林、LightGBM、XGBoost、GBDT、SVM共5种分类器中选出效果最佳分类器与递归特征消除算法结合进行数据挖掘,找出重要特征集并对其分类效果做比较与分析,最后运用Stacking模型融合方法优化模型。数据挖掘特征子集比同数量特征子集在准确率、召回率、精确率、F1值上分别提高了33.51%、14.54%、20.61%、24.04%;采用LightGBM和SVM模型融合可将F1值提高至92.6%。本文提出了一种有效的心音识别分类方法,挖掘出心音最重要的8个特征,为临床诊断提供参考。     

自适应噪声完备经验模态分解排列熵结合支持向量机的心音分类方法研究

针对心音信号非平稳性、非线性的特征,为了更直观地把心音信号的特征显示出来,提高分类识别的高效性,提出了一种自适应噪声完备经验模态分解（CEEMDAN）排列熵作为心音信号的特征向量,通过支持向量机（SVM）进行心音分类识别的方法。首先,将原始心音信号进行CEEMDAN,得到若干从高频到低频的模态分量（IMF）。其次,利用IMF分量与原始信号的相关系数、能量因子和信噪比来优选IMF做Hilbert变换,得到分量信号的瞬时频率,再计算各IMF排列熵值组成特征向量。最后,将特征向量输入SVM二分类器进行正常与异常心音信号的分类识别。对源自2016年PhysioNet/CinC挑战赛的100例心音样本进行正常与异常的分类,准确度达到87%。研究表明本文方法相比于常用的EMD和EEMD排列熵的方法准确度提高了18%～24%,可见,CEEMDAN排列熵结合SVM的方法能够有效识别正常和异常心音。     

基于新型双输入卷积神经网络的肌电模式识别

针对双臂协同运动中蕴含的运动信息量大,难以充分解读且识别率不高的问题,提出一种新型的双输入卷积神经网络（ND-CNN）模型。首先,根据双臂运动的特点,分别设计数据整理和模型输入两种策略。然后,利用两个结构相同、参数共享的特征提取层提取信号本身的特征和信号之间的差别特征。最后,利用所提取的两类特征实现双臂协同动作的识别。在自主设计的双臂实验中,将ND-CNN与其余3种先进的神经网络对比。实验结果表明,本文所提的ND-CNN模型在识别精度和可靠性上优于其他网络模型,能够对双臂肌电动作有效识别。     

Analysis of phonocardiogram signals using wavelet transform

Phonocardiograms (PCG) are recordings of the acoustic waves produced by the mechanical action of the heart. They generally consist of two kinds of acoustic vibrations: heart sounds and heart murmurs. Heart murmurs are often the first signs of pathological changes of the heart valves, and are usually found during auscultation in primary health care. Heart auscultation has been recognized for a long time as an important tool for the diagnosis of heart disease, although its accuracy is still insufficient to diagnose some heart diseases. It does not enable the analyst to obtain both qualitative and quantitative characteristics of the PCG signals. The efficiency of diagnosis can be improved considerably by using modern digital signal processing techniques. Therefore, these last can provide useful and valuable information on these signals. The aim of this study is to analyse PCG signals using wavelet transform. This analysis is based on an algorithm for the detection of heart sounds (the first and second sounds, S1 and S2) and heart murmurs using the PCG signal as the only source. The segmentation algorithm, which separates the components of the heart signal, is based on denoising by wavelet transform (DWT). This algorithm makes it possible to isolate individual sounds (S1 or S2) and murmurs. Thus, the analysis of various PCGs signals using wavelet transform can provide a wide range of statistical parameters related to the phonocardiogram signal.     

Analysis of phonocardiogram signals using wavelet transform

Phonocardiograms (PCG) are recordings of the acoustic waves produced by the mechanical action of the heart. They generally consist of two kinds of acoustic vibrations: heart sounds and heart murmurs. Heart murmurs are often the first signs of pathological changes of the heart valves, and are usually found during auscultation in primary health care. Heart auscultation has been recognized for a long time as an important tool for the diagnosis of heart disease, although its accuracy is still insufficient to diagnose some heart diseases. It does not enable the analyst to obtain both qualitative and quantitative characteristics of the PCG signals. The efficiency of diagnosis can be improved considerably by using modern digital signal processing techniques. Therefore, these last can provide useful and valuable information on these signals. The aim of this study is to analyse PCG signals using wavelet transform. This analysis is based on an algorithm for the detection of heart sounds (the first and second sounds, S1 and S2) and heart murmurs using the PCG signal as the only source. The segmentation algorithm, which separates the components of the heart signal, is based on denoising by wavelet transform (DWT). This algorithm makes it possible to isolate individual sounds (S1 or S2) and murmurs. Thus, the analysis of various PCGs signals using wavelet transform can provide a wide range of statistical parameters related to the phonocardiogram signal.     

Plasma cell granuloma of the thyroid gland mimicking carcinoma: a case report and review of the literature

Plasma cell granulomas (PCG) are rare tumor-like lesions consisting of sheets of polyclonal plasma cells admixed with numerous lymphocytes and other inflammatory cells surrounded by fibrous stroma. They usually appear in the lung, but involvement of diverse extrapulmonal sites has been described. PCGs occurring in the thyroid are very uncommon. Since 1981, only 11 cases have been described in the English literature. Here, we present the case of a 50-year-old Arabic man who noticed an enlargement of his thyroid gland during the previous 2 years, and he developed swallowing disturbances and a feeling of narrowness in the neck. A nearly total resection of the thyroid gland was made because of clinical suspicion of carcinoma. On histologic examination, PCG of the thyroid associated with Hashimoto's thyroiditis (HT) was diagnosed. This is the first case in which molecular pathological analyses for EBV and HHV8 DNA were made. As these were negative, distinct etiological features were suggested.     

Self-organizing arterial pressure pulse classification using neural networks: theoretical considerations and clinical applicability

A self-organizing classification system for the arterial pressure pulse based on the ART2 (adaptive resonance theory) network was developed. The system consists of a preprocessor and an ART2 recognition network. The preprocessor removes the arterial pressure pulse servo component signals from Finapres, detects the systolic pressure points and divides the acquired signals into minimal cardiac cycles. The ART2 network input is the minimal cardiac cycle detected by the preprocessor. The classification results can be used to assist physicians in evaluating the signs of abnormal and normal autonomic control and has shown its clinical applicability for the examination of the autonomic nervous system.     

Automatic measure of the split in the second cardiac sound by using the wavelet transform technique

This paper is concerned with the identification and automatic measure of the split in the second heart sound (S2) of the phonocardiogram signal (PCGs) for normal or pathological case. The second heart sound S2 consists of two acoustic components A2 and P2, the former is due to the closure of the aortic valve and the latter is due to the closure of the pulmonary valve. The aortic valve usually closes before the pulmonary valve, introducing a time delay known as "split". A automatic technique based on the discrete wavelet transform (DWT) and the continuous wavelet transform (CWT) is developed in this paper to measure the split of the second cardiac sound (S2) for the normal and pathological cases of the PCG signals. To quantify the splitting, the two components in S2 (i.e. A2 and P2) are identified and, the delay between the two components can be estimated. It is shown that the wavelet transform can provide best information and features of the split of S2 and the major components (A2 and P2) and consequently aid in medical diagnosis.     

Artificial neural network analysis of gustatory responses in the thalamic taste relay of the rat

We used an artificial neural network (ANN) as a model for analyzing single-neuron responses from the thalamic taste relay of rats. The network consisted of: (1) a layer of 44 input units, representing the responses of the 44 thalamic taste cells; (2) a layer of hidden units of varying numbers; and (3) a layer of four output units. We used the back-propagation algorithm to train the output units to discriminate among tastants based on inputs from the thalamic neurons. As the network became fully trained, we found that: (1) only two hidden units were necessary to provide nearly the full discriminative capacity of the network; (2) the loss of even a few of the input units that had the highest impact on hidden units caused a drastic reduction of discriminative power, implying that not all neurons contribute equally to the neural code; and (3) adding a temporal component to the input, by representing each 100-ms time bin as a separate input unit, increased the accuracy with which output units were able to identify tastants. We used data from behavioral discrimination tasks as a measure of the capacity of the network to identify stimuli correctly. A network with two hidden units was about as effective as an across-pattern analysis in accounting for the rat's discriminative ability.     

A neural network method for identification of RNA-interacting residues in protein

Identification of the most putative RNA-interacting residues in protein is an important and challenging problem in a field of molecular recognition. Structural analysis of protein-RNA complexes reveals a strong correlation between interaction residues and their structure. Building on this viewpoint, we have developed a neural network predictor to correctly identify residues involved in protein-RNA interactions from protein sequence and its structural information. The system has been exhaustedly cross-validated with various strategies differing in input encoding, amount of input information, and network architectures. In addition, we have evaluated performance among functional subsets of complexes. Finally, to reflect the properties of protein-RNA complexes in our dataset, two kinds of post-processing method are adopted. The experimental result shows that our system yields not-trivial performance although the residues in interaction sites are too scarce.     

Neural correlates of gesture processing across human development

Co-speech gesture facilitates learning to a greater degree in children than in adults, suggesting that the mechanisms underlying the processing of co-speech gesture differ as a function of development. We suggest that this may be partially due to children's lack of experience producing gesture, leading to differences in the recruitment of sensorimotor networks when comparing adults to children. Here, we investigated the neural substrates of gesture processing in a cross-sectional sample of 5-, 7.5-, and 10-year-old children and adults and focused on relative recruitment of a sensorimotor system that included the precentral gyrus (PCG) and the posterior middle temporal gyrus (pMTG). Children and adults were presented with videos in which communication occurred through different combinations of speech and gesture during a functional magnetic resonance imaging (fMRI) session. Results demonstrated that the PCG and pMTG were recruited to different extents in the two populations. We interpret these novel findings as supporting the idea that gesture perception (pMTG) is affected by a history of gesture production (PCG), revealing the importance of considering gesture processing as a sensorimotor process.     

Modelling the time-keeping function of the central pattern generator for locomotion using artificial sequential neural network

The paper investigates the ability of a sequential neural network to model the time-keeping function (fundamental frequency oscillation) of a central pattern generator for locomotion. The intention is not to strive for biological fidelity, but rather to ensure that the network obeys the organisational and operational principles of central pattern generators developed through empirical research. The timing function serves to produce the underlying locomotor rhythm which can be transformed by nonlinear static shaping functions to construct the necessary locomotor activation patterns. Using two levels of tonic activations in the form of a step increase, a network consisting of nine processing units was successfully trained to output both sine and cosine waveforms, whose frequencies were modified in response to the level of input activation. The network's ability to generalise was demonstrated by appropriately scaling the frequency of oscillation in response to a range of input amplitudes, both within and outside the values on which it was trained. A notable and fortunate result was the model's failure to oscillate in the absence of input activation, which is a necessary property of the CPG model. It was further demonstrated that the oscillation frequency of the output waveforms exhibited both a high temporal stability and a very low sensitivity to input noise. The results indicate that the sequential neural network is a suitable candidate to model the time-keeping functions of the central pattern generator for locomotion.     

Detection of systolic ejection click using time growing neural network

In this paper, we present a novel neural network for classification of short-duration heart sounds: the time growing neural network (TGNN). The input to the network is the spectral power in adjacent frequency bands as computed in time windows of growing length. Children with heart systolic ejection click (SEC) and normal children are the two groups subjected to analysis. The performance of the TGNN is compared to that of a time delay neural network (TDNN) and a multi-layer perceptron (MLP), using training and test datasets of similar sizes with a total of 614 normal and abnormal cardiac cycles. From the test dataset, the classification rate/sensitivity is found to be 97.0%/98.1% for the TGNN, 85.1%/76.4% for the TDNN, and 92.7%/85.7% for the MLP. The results show that the TGNN performs better than do TDNN and MLP when frequency band power is used as classifier input. The performance of TGNN is also found to exhibit better immunity to noise.