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。     

基于迁移学习的深层卷积神经网络心电信号疲劳分类

传统的心电疲劳分类方法虽然能有效地识别疲劳状态,但需要采集较长时间的信号,不能达到疲劳状态的实时监测。本文设计一种深层卷积神经网络模型用于评估操作员疲劳状态,对操作员的短时心电信号进行疲劳状态的自动分类。首先,提出一种将心电信号转化为图像的方法,将采集到的心电信号转化成二维图像,即将心电信号直接映射到二维空间转换成时域图片信息。然后,将图片送入深层卷积神经网络模型中去训练,实现对操作员疲劳状态的分类。本文方法降低了模型的复杂性,减少了模型的参数,同时训练的数据不需要经过类似噪声滤波、特征提取等任何预处理步骤。结果表明该模型能自动从心电信号中提取有效特征,实现对操作员非疲劳和疲劳两种状态的正确分类,分类准确率达到97.36%。     

A deep learning-based algorithm for detection of cortical arousal during sleep

Study ObjectivesThe frequency of cortical arousals is an indicator of sleep quality. Additionally, cortical arousals are used to identify hypopneic events. However, it is inconvenient to record electroencephalogram (EEG) data during home sleep testing. Fortunately, most cortical arousal events are associated with autonomic nervous system activity that could be observed on an electrocardiography (ECG) signal. ECG data have lower noise and are easier to record at home than EEG. In this study, we developed a deep learning-based cortical arousal detection algorithm that uses a single-lead ECG to detect arousal during sleep.MethodsThis study included 1,547 polysomnography records that met study inclusion criteria and were selected from the Multi-Ethnic Study of Atherosclerosis database. We developed an end-to-end deep learning model consisting of convolutional neural networks and recurrent neural networks which: (1) accepted varying length physiological data; (2) directly extracted features from the raw ECG signal; (3) captured long-range dependencies in the physiological data; and (4) produced arousal probability in 1-s resolution.ResultsWe evaluated the model on a test set (n = 311). The model achieved a gross area under precision-recall curve score of 0.62 and a gross area under receiver operating characteristic curve score of 0.93.ConclusionThis study demonstrated the end-to-end deep learning approach with a single-lead ECG has the potential to be used to accurately detect arousals in home sleep tests.     

Artificial neural network-based classification of body movements in ambulatory ECG signal

Abstract

Ambulatory ECG monitoring provides electrical activity of the heart when a person is involved in doing normal routine activities. Thus, the recorded ECG signal consists of cardiac signal along with motion artifacts introduced due to a person’s body movements during routine activities. Detection of motion artifacts due to different physical activities might help in further cardiac diagnosis. Ambulatory ECG signal analysis for detection of various motion artifacts using adaptive filtering approach is addressed in this paper. We have used BIOPAC MP 36 system for acquiring ECG signal. The ECG signals of five healthy subjects (aged between 22–30 years) were recorded while the person performed various body movements like up and down movement of the left hand, up and down movement of the right hand, waist twisting movement while standing and change from sitting down on a chair to standing up movement in lead I configuration. An adaptive filter-based approach has been used to extract the motion artifact component from the ambulatory ECG signal. The features of motion artifact signal, extracted using Gabor transform, have been used to train the artificial neural network (ANN) for classifying body movements.     

基于1D-ECNN的短时心电信号疲劳分类

提出一维双卷积神经网络（1D-ECNN），基于采集的心电信号检测操作员的疲劳状态。1D-ECNN包括4 个卷积 层、2个最大池化层、1个全连接层和1个softmax输出层。本研究仅使用较少的卷积核数量，这将减少模型参数的数量，降 低模型的复杂程度，提高模型训练的速度，同时避免传统方法中复杂的特征提取过程或特征选择过程。将心电信号分成 时间长度为1 s的样本，送入1D-ECNN，基于短时心电信号进行操作员疲劳状态分类。仿真结果表明，本文方法的平均分 类准确率高达95.72%，能够实时准确地检测操作员的疲劳状态。此外，可以较好地消除个体差异性的影响。     

Deep learning‐based fully automatic segmentation of wrist cartilage in MR images

The study objective was to investigate the performance of a dedicated convolutional neural network (CNN) optimized for wrist cartilage segmentation from 2D MR images. CNN utilized a planar architecture and patch‐based (PB) training approach that ensured optimal performance in the presence of a limited amount of training data. The CNN was trained and validated in 20 multi‐slice MRI datasets acquired with two different coils in 11 subjects (healthy volunteers and patients). The validation included a comparison with the alternative state‐of‐the‐art CNN methods for the segmentation of joints from MR images and the ground‐truth manual segmentation. When trained on the limited training data, the CNN outperformed significantly image‐based and PB‐U‐Net networks. Our PB‐CNN also demonstrated a good agreement with manual segmentation (Sørensen–Dice similarity coefficient [DSC] = 0.81) in the representative (central coronal) slices with a large amount of cartilage tissue. Reduced performance of the network for slices with a very limited amount of cartilage tissue suggests the need for fully 3D convolutional networks to provide uniform performance across the joint. The study also assessed inter‐ and intra‐observer variability of the manual wrist cartilage segmentation (DSC = 0.78‐0.88 and 0.9, respectively). The proposed deep learning‐based segmentation of the wrist cartilage from MRI could facilitate research of novel imaging markers of wrist osteoarthritis to characterize its progression and response to therapy.     

Predicting Deletion of Chromosomal Arms 1p/19q in Low-Grade Gliomas from MR Images Using Machine Intelligence

Several studies have linked codeletion of chromosome arms 1p/19q in low-grade gliomas (LGG) with positive response to treatment and longer progression-free survival. Hence, predicting 1p/19q status is crucial for effective treatment planning of LGG. In this study, we predict the 1p/19q status from MR images using convolutional neural networks (CNN), which could be a non-invasive alternative to surgical biopsy and histopathological analysis. Our method consists of three main steps: image registration, tumor segmentation, and classification of 1p/19q status using CNN. We included a total of 159 LGG with 3 image slices each who had biopsy-proven 1p/19q status (57 non-deleted and 102 codeleted) and preoperative postcontrast-T1 (T1C) and T2 images. We divided our data into training, validation, and test sets. The training data was balanced for equal class probability and was then augmented with iterations of random translational shift, rotation, and horizontal and vertical flips to increase the size of the training set. We shuffled and augmented the training data to counter overfitting in each epoch. Finally, we evaluated several configurations of a multi-scale CNN architecture until training and validation accuracies became consistent. The results of the best performing configuration on the unseen test set were 93.3% (sensitivity), 82.22% (specificity), and 87.7% (accuracy). Multi-scale CNN with their self-learning capability provides promising results for predicting 1p/19q status non-invasively based on T1C and T2 images. Predicting 1p/19q status non-invasively from MR images would allow selecting effective treatment strategies for LGG patients without the need for surgical biopsy.     

基于卷积神经网络和长短时记忆网络的心肌梗死检测

为了进一步提升心肌梗死的诊断效果,提出一种基于卷积神经网络和长短时记忆网络的心肌梗死检测方法,用于准确地从心电图信号中检测心肌梗死。具体来讲,提出3种分别基于卷积神经网络、卷积神经网络结合长短时记忆网络以及它们的集成模型的检测方法,以期从心电信号中检测心肌梗死和正常搏动。此外,采用数据重采样方法,即合成少数类过采样方法和Tomek Link解决数据集不平衡问题。最终与其他方法的实验结果相比,经过数据重采样的集成卷积神经网络模型的结果取得了明显优势,证明提出方法的有效性。     

基于多分支融合神经网络的心电图导联重构方法

设计一种新型的多分支信息融合神经网络结构,利用已知的I,Ⅱ,V2 3个导联心电信号来重构其它导联心电信号。基于卷积神经网络结构提取多个导联的特征然后进行线性相加融合,采用一种改进的双向长短期记忆网络结构来获得与时序相关的信息,从而实现心电图导联重构。使用Physikalisch Technische Bundesanstalt(PTB)数据库进行验证,导联重构方法具有0.944 4的相关系数和0.320 3的均方根误差,说明新型神经网络结构可以有效地实现心电图导联重构。     

An ECG simulator with a novel ECG profile for physiological signals

Abstract

This work introduces a low-cost open-source electrocardiography (ECG) simulator comprising both MATLAB software for signal generation and a dedicated circuit board for signal output via a commercial sound card. Synthetic, rate-dependent ECG simulation is based on third-order polynomials that are calculated in sections for the main waves and spikes, respectively. Besides the heart rate, the output profile is fully adjustable with respect to Einthoven lead signals I–III, the amplitudes of the individual ECG waves and spikes, as well as the constitution and intensity of common distortions. The underlying coefficients for the synthetic ECG profile are obtained experimentally by analysing recordings of 22 healthy individuals with heart rates in the range of 40–180 bpm. Eight of these recordings are selected to determine the coefficients for the polynomials (training set) while the remaining 14 serve as test set to evaluate their applicability and accuracy. Thereby, a mean correlation of 98.57% is found which is superior in comparison with a widely accepted rate-dependent ECG profile that is generated from square root and linear terms (correlation score: 91.46%). Although other use-cases are feasible, the focus of this work is the development of an ECG simulator for academic research and university education. Both the MATLAB source code and the circuit layout files are available in the online supplement stimulating further work on this topic.     

Automatically extracting information needs from complex clinical questions

ObjectiveClinicians pose complex clinical questions when seeing patients, and identifying the answers to those questions in a timely manner helps improve the quality of patient care. We report here on two natural language processing models, namely, automatic topic assignment and keyword identification, that together automatically and effectively extract information needs from ad hoc clinical questions. Our study is motivated in the context of developing the larger clinical question answering system AskHERMES (Help clinicians to Extract and aRrticulate Multimedia information for answering clinical quEstionS).Design and measurementsWe developed supervised machine-learning systems to automatically assign predefined general categories (e.g. etiology, procedure, and diagnosis) to a question. We also explored both supervised and unsupervised systems to automatically identify keywords that capture the main content of the question.ResultsWe evaluated our systems on 4654 annotated clinical questions that were collected in practice. We achieved an F1 score of 76.0% for the task of general topic classification and 58.0% for keyword extraction. Our systems have been implemented into the larger question answering system AskHERMES. Our error analyses suggested that inconsistent annotation in our training data have hurt both question analysis tasks.ConclusionOur systems, available at http://www.askhermes.org, can automatically extract information needs from both short (the number of word tokens <20) and long questions (the number of word tokens >20), and from both well-structured and ill-formed questions. We speculate that the performance of general topic classification and keyword extraction can be further improved if consistently annotated data are made available.     

基于空洞卷积神经网络的心律失常分类算法

本文提出了一种基于卷积网络的心电信号分类算法,设计了空洞卷积池化金字塔模块,通过不同尺寸的空洞卷积提取信息,再将各通道的信息聚合,在增强网络的特征提取能力的同时可以降低参数量。本文聚焦于窦性心律、房性早搏、心动过速以及心动过缓4种分类,使用的心电图数据集来自医院的实测数据,数据集包含75000名不同检测者的心电记录。经过测试,本文提出的模型在该数据集上取得了0.89的F1值,另外在CinC2017数据集上也达到了0.87的F1值。实验结果表明该分类算法具有优秀的特征提取和分类能力,在心电信号的实时分类中具备应用前景。     

Comparison of 18F SPECT with PET in myocardial imaging: A realistic thorax-cardiac phantom study

Background

Cardiac syndrome X is defined by a typical angina pectoris with normal or near normal (stenosis <40%) coronary angiogram with or without electrocardiogram (ECG) change or atypical angina pectoris with normal or near normal coronary angiogram plus a positive none-invasive test (exercise tolerance test or myocardial perfusion scan) with or without ECG change. Studies with myocardial perfusion imaging on this syndrome have indicated some abnormal perfusion scan. We evaluated the role of myocardial perfusion imaging (MPI) and also the severity and extent of perfusion abnormality using Tc-99m MIBI Single Photon Emission Computed Tomography (SPECT) in these patients.

Methods

The study group consisted of 36 patients with cardiac syndrome X. The semiquantitative perfusion analysis was performed using exercise Tc-99m MIBI SPECT. The MPI results were analyzed by the number, location and severity of perfusion defects.

Results

Abnormal perfusion defects were detected in 13 (36.10%) cases, while the remaining 23 (63.90%) had normal cardiac imaging. Five of 13 (38.4%) abnormal studies showed multiple perfusion defects. The defects were localized in the apex in 3, apical segments in 4, midventricular segments in 12 and basal segments in 6 cases. Fourteen (56%) of all abnormal segments revealed mild, 7(28%) moderate and 4 (16%) severe reduction of tracer uptake. No fixed defects were identified. The vessel territories were approximately the same in all subjects. The Exercise treadmill test (ETT) was positive in 25(69%) and negative in 11(30%) patients. There was no consistent pattern as related to the extent of MPI defects or exercise test results.

Conclusion

Our study suggests that multiple perfusion abnormalities with different levels of severity are common in cardiac syndrome X, with more than 30 % of these patients having at least one abnormal perfusion segment. Our findings suggest that in these patients microvascular angina is probably more common than is generally believed.     

Detection of obstructive sleep apnea in pediatric subjects using surface lead electrocardiogram features

STUDY OBJECTIVES: To investigate the feasibility of detecting obstructive sleep apnea (OSA) in children using an automated classification system based on analysis of overnight electrocardiogram (ECG) recordings. DESIGN: Retrospective observational study. SETTING: A pediatric sleep clinic. PARTICIPANTS: Fifty children underwent full overnight polysomnography. INTERVENTION: N/A. MEASUREMENTS AND RESULTS: Expert polysomnography scoring was performed. The datasets were divided into a training set of 25 subjects (11 normal, 14 with OSA) and a withheld test set of 25 subjects (11 normal, 14 with OSA). Features, calculated from the ECG of the 25 training datasets, were empirically chosen to train a modified quadratic discriminant analysis classification system. The selected configuration used a segment length of 60 seconds and processed mean, SD, power spectral density, and serial correlation measures to classify segments as apneic or normal. By combining per-segment classifications and using receiver-operator characteristic analysis, a per-subject classifier was obtained that had a sensitivity of 85.7%, specificity of 90.9%, and accuracy of 88% on the training datasets. The same decision threshold was applied to the withheld datasets and yielded a sensitivity of 85.7%, specificity of 81.8%, and accuracy of 84%. The positive and negative predictive values were 85.7% and 81.8%, respectively, on the test dataset. CONCLUSIONS: The ability to correctly identify 12 out of 14 cases of OSA (with the 2 false negatives arising from subjects with an apnea-hypopnea index less than 10) indicates that the automated apnea classification system outlined may have clinical utility in pediatric patients.     

Assessment of obstructive sleep apnea-related sleep fragmentation utilizing deep learning-based sleep staging from photoplethysmography

Study ObjectivesTo assess the relationship between obstructive sleep apnea (OSA) severity and sleep fragmentation, accurate differentiation between sleep and wakefulness is needed. Sleep staging is usually performed manually using electroencephalography (EEG). This is time-consuming due to complexity of EEG setup and the amount of work in manual scoring. In this study, we aimed to develop an automated deep learning-based solution to assess OSA-related sleep fragmentation based on photoplethysmography (PPG) signal.MethodsA combination of convolutional and recurrent neural networks was used for PPG-based sleep staging. The models were trained using two large clinical datasets from Israel (n = 2149) and Australia (n = 877) and tested separately on three-class (wake/NREM/REM), four-class (wake/N1 + N2/N3/REM), and five-class (wake/N1/N2/N3/REM) classification. The relationship between OSA severity categories and sleep fragmentation was assessed using survival analysis of mean continuous sleep. Overlapping PPG epochs were applied to artificially obtain denser hypnograms for better identification of fragmented sleep.ResultsAutomatic PPG-based sleep staging achieved an accuracy of 83.3% on three-class, 74.1% on four-class, and 68.7% on five-class models. The hazard ratios for decreased mean continuous sleep compared to the non-OSA group obtained with Cox proportional hazards models with 5-s epoch-to-epoch intervals were 1.70, 3.30, and 8.11 for mild, moderate, and severe OSA, respectively. With EEG-based hypnograms scored manually with conventional 30-s epoch-to-epoch intervals, the corresponding hazard ratios were 1.18, 1.78, and 2.90.ConclusionsPPG-based automatic sleep staging can be used to differentiate between OSA severity categories based on sleep continuity. The differences between the OSA severity categories become more apparent when a shorter epoch-to-epoch interval is used.     

Convolutional Neural Network to Segment Laminae on 3D Ultrasound Spinal Images to Assist Cobb Angle Measurement

A recent innovation in scoliosis monitoring is the use of ultrasonography, which provides true 3D information in one scan and does not emit ionizing radiation. Measuring the severity of scoliosis on ultrasonographs requires identifying lamina pairs on the most tilted vertebrae, which is difficult and time-consuming. To expedite and automate measurement steps, this paper detailed an automatic convolutional neural network-based algorithm for identifying the laminae on 3D ultrasonographs. The predicted laminae were manually paired to measure the lateral spinal curvature on the coronal view, called the Cobb angle. In total, 130 spinal ultrasonographs of adolescents with idiopathic scoliosis recruited from a scoliosis clinic were selected, with 70 for training and 60 for testing. Data augmentation increased the effective training set size to 140 ultrasonographs. Semi-automatic Cobb measurements were compared to manual measurements on the same ultrasonographs. The semi-automatic measurements demonstrated good inter-method reliability (ICC_3,1 = 0.87) and performed better on thoracic (ICC_3,1 = 0.91) than lumbar curves (ICC_3,1 = 0.81). The mean absolute difference and standard deviation between semi-automatic and manual was 3.6° ± 3.0°. In conclusion, the semi-automatic method to measure the Cobb angle on ultrasonographs is feasible and accurate. This is the first algorithm that automates steps of Cobb angle measurement on ultrasonographs.

     

Respiratory rate estimation during triage of children in hospitals

Abstract

Accurate assessment of a child’s health is critical for appropriate allocation of medical resources and timely delivery of healthcare in Emergency Departments. The accurate measurement of vital signs is a key step in the determination of the severity of illness and respiratory rate is currently the most difficult vital sign to measure accurately. Several previous studies have attempted to extract respiratory rate from photoplethysmogram (PPG) recordings. However, the majority have been conducted in controlled settings using PPG recordings from healthy subjects. In many studies, manual selection of clean sections of PPG recordings was undertaken before assessing the accuracy of the signal processing algorithms developed. Such selection procedures are not appropriate in clinical settings. A major limitation of AR modelling, previously applied to respiratory rate estimation, is an appropriate selection of model order. This study developed a novel algorithm that automatically estimates respiratory rate from a median spectrum constructed applying multiple AR models to processed PPG segments acquired with pulse oximetry using a finger probe. Good-quality sections were identified using a dynamic template-matching technique to assess PPG signal quality. The algorithm was validated on 205 children presenting to the Emergency Department at the John Radcliffe Hospital, Oxford, UK, with reference respiratory rates up to 50 breaths per minute estimated by paediatric nurses. At the time of writing, the authors are not aware of any other study that has validated respiratory rate estimation using data collected from over 200 children in hospitals during routine triage.     

Efficiency Improvement in a Busy Radiology Practice: Determination of Musculoskeletal Magnetic Resonance Imaging Protocol Using Deep-Learning Convolutional Neural Networks

The purposes of this study are to evaluate the feasibility of protocol determination with a convolutional neural networks (CNN) classifier based on short-text classification and to evaluate the agreements by comparing protocols determined by CNN with those determined by musculoskeletal radiologists. Following institutional review board approval, the database of a hospital information system (HIS) was queried for lists of MRI examinations, referring department, patient age, and patient gender. These were exported to a local workstation for analyses: 5258 and 1018 consecutive musculoskeletal MRI examinations were used for the training and test datasets, respectively. The subjects for pre-processing were routine or tumor protocols and the contents were word combinations of the referring department, region, contrast media (or not), gender, and age. The CNN Embedded vector classifier was used with Word2Vec Google news vectors. The test set was tested with each classification model and results were output as routine or tumor protocols. The CNN determinations were evaluated using the receiver operating characteristic (ROC) curves. The accuracies were evaluated by a radiologist-confirmed protocol as the reference protocols. The optimal cut-off values for protocol determination between routine protocols and tumor protocols was 0.5067 with a sensitivity of 92.10%, a specificity of 95.76%, and an area under curve (AUC) of 0.977. The overall accuracy was 94.2% for the ConvNet model. All MRI protocols were correct in the pelvic bone, upper arm, wrist, and lower leg MRIs. Deep-learning-based convolutional neural networks were clinically utilized to determine musculoskeletal MRI protocols. CNN-based text learning and applications could be extended to other radiologic tasks besides image interpretations, improving the work performance of the radiologist.     

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。