Blood Pressure Estimation Based on PPG and ECG Signals Using Knowledge Distillation.

OBJECTIVE: Easy access bio-signals are useful for alleviating the shortcomings and difficulties associated with cuff-based and invasive blood pressure (BP) measurement techniques. This study proposes a deep learning model, trained using knowledge distillation, based on photoplethysmographic (PPG) and electrocardiogram (ECG) signals to estimate systolic and diastolic blood pressures. METHODS: The estimation model comprises convolutional layers followed by one bidirectional recurrent layer and attention layers. The training approach involves knowledge distillation, where a smaller model (student model) is trained by leveraging information from a larger model (teacher model). RESULTS: The proposed multistage model was evaluated on 1205 subjects from Medical Information Mart for Intensive Care (MIMIC) III database using the Association for the Advancement of Medical Instrumentation (AAMI) and the standards of the British Hypertension Society (BHS). The results revealed that our model performance achieved grade A in estimating both systolic blood pressure (SBP) and diastolic blood pressure (DBP) and met the requirements of the AAMI standard. After training with knowledge distillation (KD), the model achieved a mean absolute error and standard deviation of 2.94 ± 5.61 mmHg for SBP and 2.02 ± 3.60 mmHg for DBP. CONCLUSION: Our results demonstrate the benefits of the knowledge distillation training method in reducing the number of parameters and improving the predictive accuracy of the blood pressure regression model.

An Arrhythmia Classification Model Based on Vision Transformer with Deformable Attention.

The electrocardiogram (ECG) is a highly effective non-invasive tool for monitoring heart activity and diagnosing cardiovascular diseases (CVDs). Automatic detection of arrhythmia based on ECG plays a critical role in the early prevention and diagnosis of CVDs. In recent years, numerous studies have focused on using deep learning methods to address arrhythmia classification problems. However, the transformer-based neural network in current research still has a limited performance in detecting arrhythmias for the multi-lead ECG. In this study, we propose an end-to-end multi-label arrhythmia classification model for the 12-lead ECG with varied-length recordings. Our model, called CNN-DVIT, is based on a combination of convolutional neural networks (CNNs) with depthwise separable convolution, and a vision transformer structure with deformable attention. Specifically, we introduce the spatial pyramid pooling layer to accept varied-length ECG signals. Experimental results show that our model achieved an F1 score of 82.9% in CPSC-2018. Notably, our CNN-DVIT outperforms the latest transformer-based ECG classification algorithms. Furthermore, ablation experiments reveal that the deformable multi-head attention and depthwise separable convolution are both efficient in extracting features from multi-lead ECG signals for diagnosis. The CNN-DVIT achieved good performance for the automatic arrhythmia detection of ECG signals. This indicates that our research can assist doctors in clinical ECG analysis, providing important support for the diagnosis of arrhythmia and contributing to the development of computer-aided diagnosis technology.

A lightweight U-net for ECG denoising using knowledge distillation.

Objective.Electrocardiogram (ECG) signals are easily polluted by various noises which are likely to have adverse effects on subsequent interpretations. Research on model lightweighting can promote the practical application of deep learning-based ECG denoising methods in real-time processing.Approach.Firstly, grouped convolution and conventional convolution are combined to replace the continuous conventional convolution in the model, and the depthwise convolution with stride is used to compress the feature map in the encoder modules. Secondly, additional identity connections and a local maximum and minimum enhancement module are designed, which can retain the detailed information and characteristic waveform in the ECG waveform while effectively denoising. Finally, we develop knowledge distillation in the experiments, which further improves the ECG denoising performance without increasing the model complexity. The ground-truth ECG is from The China Physiological Signal Challenge (CPSC) 2018, and the noise signal is from the MIT-BIH Noise Stress Test Database (NSTDB). We evaluate denoising performance using the signal-to-noise ratio (SNR), the root mean square error (RMSE) and the Pearson correlation coefficient (P). We use the floating point of operations (FLOPs) and parameters to calculate computational complexity.Main Results.Different data generation processes are used to conduct experiments: group 1, group 2 and group 3. The results show that the proposed model (ULde-net) can improve SNRs by 10.30 dB, 12.16 dB and 12.61 dB; reduce RMSEs by 9.88 × 10-2, 20.63 × 10-2and 15.25 × 10-2; and increasePs by 14.77 × 10-2, 27.74 × 10-2and 21.32 × 10-2. Moreover, the denoising performance after knowledge distillation is further improved. The ULde-net has parameters of 6.9 K and FLOPs of 6.6 M, which are much smaller than the compared models.Significance.We designed a lightweight model, but also retain adequate ECG denoising performance. We believe that this method can be successfully applied to practical applications under time or memory limits.

Detection of arrhythmia in 12-lead varied-length ECG using multi-branch signal fusion network.

Objective.Automatic detection of arrhythmia based on electrocardiogram (ECG) plays a critical role in early prevention and diagnosis of cardiovascular diseases. With the increase in widely available digital ECG data and the development of deep learning, multi-class arrhythmia classification based on automatic feature extraction of ECG has become increasingly attractive. However, the majority of studies cannot accept varied-length ECG signals and have limited performance in detecting multi-class arrhythmias.Approach.In this study, we propose a multi-branch signal fusion network (MBSF-Net) for multi-label classification of arrhythmia in 12-lead varied-length ECG. Our model utilizes the complementary power between different structures, which include Inception with depthwise separable convolution (DWS-Inception), spatial pyramid pooling (SPP) Layer, and multi-scale fusion Resnet (MSF-Resnet). The proposed method can extract features from each lead of 12-lead ECG recordings separately and then effectively fuse the features of each lead by integrating multiple convolution kernels with different receptive fields, which can achieve the information of complementation between different angles of the ECG signal. In particular, our model can accept 12-lead ECG signals of arbitrary length.Main results.The experimental results show that our model achieved an overall classification F1 score of 83.8% in the 12-lead ECG data of CPSC-2018. In addition, the F1 score of the MBSF-Net performed best among the MBF-Nets which are removed the SPP layer from MBSF-Net. In comparison with the latest ECG classification algorithms, the proposed model can be applied in varied-length signals and has an excellent performance, which not only can fully retain the integrity of the original signals, but also eliminates the cropping/padding signal beforehand when dealing with varied-length signal database.Significance.MBSF-Net provides an end-to-end multi-label classification model with outperfom performance, which allows detection of disease in varied-length signals without any additional cropping/padding. Moreover, our research is beneficial to the development of computer-aided diagnosis.

Supraventricular ectopic beats and ventricular ectopic beats detection based on improved U-net.

Objective.Supraventricular ectopic beats (SVEB) or ventricular ectopic beats (VEB) are common arrhythmia with uncertain occurrence and morphological diversity, so realizing their automatic localization is of great significance in clinical diagnosis.Methods.We propose a modified U-net network: USV-net, it can simultaneously realize the automatic positioning of VEB and SVEB. The improvement consists of three parts: firstly, we reconstruct part of the convolutional layer in U-net using group convolution to reduce the expression of redundant features. Secondly, a plug-and-play multi-scale 2D deformable convolution module is designed to extract cross-channel features of different scales. Thirdly, in addition to conventional output of U-net, we also compress and output the bottom feature map of U-net, the dual-output is trained through Dice-loss to take into account the learning of high/low resolution features of the model. We used the MIT-BIH arrhythmia database for patient-specific training, and used Sensitivity, Positive prediction rate and F1-scores to evaluate the effectiveness of our method.Main Result.The F1-scores of SVEB and VEB achieve the best results compared with other studies in different testing records. It is worth noting that the F1-scores of SVEB and VEB reached 81.3 and 95.4 in the 24 testing records. Moreover, our method is also at the forefront in Sensitivity and Positive prediction rate.Significance.The method proposed in this paper has great potential in the detection of SVEB and VEB. We anticipate efficiency and accuracy of clinical detection of ectopic beats would be improved.

Inter-patient arrhythmia identification method with RR-intervals and convolutional neural networks.

Objective. The arrhythmia identification method based on the U-net has the potential for fast application. The RR-intervals have been proven to improve the performance of single-heartbeat identification methods. However, because both the heartbeats number and location in the input of the U-net are unfixed, the approach based on the U-net cannot use RR-intervals directly. To solve this problem, we proposed a novel method. The proposed method also can identify heartbeats of four classes, including non-ectopic (N), supraventricular ectopic beat (SVEB), ventricular ectopic beat (VEB), and fusion beat (F).Approach. Our method consists of the pre-processing and the two-stage identification framework. In the pre-processing part, we filtered input signals with a band-pass filter and created the auxiliary waveforms by RR-intervals. In the first stage of the framework, we designed a network to handle input signals and auxiliary waveforms. We proposed a masking operation to separate the input signal into two signals according to the result of the network. The first signal contains heartbeats of SVEB and VEB. The second signal includes heartbeats of N and F. The second stage consists of two networks and can further identify the heartbeats of SVEB, VEB, N, and F from these two signals.Main result. We validated our method on the MIT-BIH arrhythmia database with the inter-patient model. For classes N, SVEB, VEB, and F, our approach achieved F1 scores of 98.26, 68.61, 95.99, and 47.75, respectively.Significance. Our method not only can effectively utilize RR intervals but also can identify multiple arrhythmias.

Two-stage ECG signal denoising based on deep convolutional network.

Background.An electrocardiogram (ECG) is an effective and non-invasive indicator for the detection and prevention of arrhythmia. ECG signals are susceptible to noise contamination, which can lead to errors in ECG interpretation. Therefore, ECG pretreatment is important for accurate analysis.Methods.The ECG data used are from CPSC2018, and the noise signal is from MIT-BIH Noise Stress Test Database. In the experiment, the signal-to-noise ratio (SNR), the root mean square error (RMSE), and the correlation coefficientPare used to evaluate the performance of the network. The method proposed is divided into two stages. In the first stage, a Ude-net model is designed for ECG signal denoising to eliminate noise. The DR-net model in the second stage is used to reconstruct the ECG signal and to correct the waveform distortion caused by noise removal in the first stage. In this paper, the Ude-net and the DR-net are constructed by the convolution method to achieve end-to-end mapping from noisy ECG signals to clean ECG signals.Result.In SNR, RMSE andPindicators, Ude-net + DR-net proposed in this paper can achieve the best performance compared with the other five schemes (FCN, U-net etc). In the three data sets, SNR can be increased by 11.61 dB, 13.71 dB and 14.40 dB and RMSE can be reduced by 10.46 × 10-2, 21.55 × 10-2and 15.98 × 10-2.Conclusions.Despite the contradictory results, the proposed two-stages method can achieve both the elimination of noise and the preservation of effective details to a large extent of the signals. The proposed method has good application prospects in clinical practice.

A novel method to reduce false alarms in ECG diagnostic systems: capture and quantification of noisy signals.

Objective. Muscle artifacts (MA) and electrode motion artifacts (EMA) in electrocardiogram (ECG) signals lead to a large number of false alarms from cardiac diagnostic systems. To reduce false alarms, it is necessary to improve the performance of the diagnostic algorithm in noisy environments by removing excessively noisy signals. However, existing methods focus on signal quality assessment and contain too many artificial features. Here, we present a novel method to flexibly eliminate noisy signals without any artificial features.Approach. Our method contains an improved lightweight deep neural network (DNN) to capture the signal portions damaged by EMA and MA, uses the sample entropy to quantize noisy portions, and discards those portions that exceed a defined threshold. Our method was tested in conjunction with Pan-Tompkins (PT), Filter Bank (FB), and 'UNSW' R-peak detection algorithms along with two heartbeat classification algorithms on datasets synthesized from the MIT-BIH Noise Stress Test Database, the China Physiological Signal Challenge 2018 Database and the MIT-BIH Arrhythmia Database.Main results. For PT R-peak detection algorithms, the sensitivity (Se) increased noticeably from 89.01% to 99.42% in the synthesized datasets with a signal-to-noise ratio of 6 dB. With the same datasets, the Se of the FB algorithm increased about 9.29%, and a 3.64% increase occurred in the Se of the 'UNSW' algorithm. For heartbeat classification algorithms, the overall F1-score increased about 6% in the synthesized one-heartbeat datasets. It is the first study to utilize a DNN to capture noisy segments of the ECG signal.Significance. Too many false alarms can cause alarm fatigue. Our method can be utilized as the preprocessing before signal analysis, thereby reducing false alarms from the ECG diagnostic systems.