Very‐large‐scale integration implementation of a convolutional neural network accelerator for abnormal heartbeat detection
2020; Institution of Engineering and Technology; Volume: 56; Issue: 7 Linguagem: Inglês
10.1049/el.2019.3752
ISSN1350-911X
Autores Tópico(s)Phonocardiography and Auscultation Techniques
ResumoElectronics LettersVolume 56, Issue 7 p. 330-331 Circuits and systemsFree Access Very-large-scale integration implementation of a convolutional neural network accelerator for abnormal heartbeat detection Y.-H. Chen, Corresponding Author Y.-H. Chen chenyh@mail.cgu.edu.tw orcid.org/0000-0001-5651-7584 Department of Electronics Engineering, Chang Gung University, Taiwan Y.-H. Chen: Also with Department of Radiation Oncology, Institute for Radiological Research, Chang Gung Memorial Hospital-LinKou, TaiwanSearch for more papers by this authorY. Juan, Y. Juan Department of Electronics Engineering, Chang Gung University, TaiwanSearch for more papers by this author Y.-H. Chen, Corresponding Author Y.-H. Chen chenyh@mail.cgu.edu.tw orcid.org/0000-0001-5651-7584 Department of Electronics Engineering, Chang Gung University, Taiwan Y.-H. Chen: Also with Department of Radiation Oncology, Institute for Radiological Research, Chang Gung Memorial Hospital-LinKou, TaiwanSearch for more papers by this authorY. Juan, Y. Juan Department of Electronics Engineering, Chang Gung University, TaiwanSearch for more papers by this author First published: 01 March 2020 https://doi.org/10.1049/el.2019.3752Citations: 4 AboutSectionsPDF ToolsRequest permissionExport citationAdd to favoritesTrack citation ShareShare Give accessShare full text accessShare full-text accessPlease review our Terms and Conditions of Use and check box below to share full-text version of article.I have read and accept the Wiley Online Library Terms and Conditions of UseShareable LinkUse the link below to share a full-text version of this article with your friends and colleagues. Learn more.Copy URL Share a linkShare onFacebookTwitterLinkedInRedditWechat Abstract In this study, a very-large-scale integration implementation of a convolutional neural network (CNN) inference for abnormal heartbeat detection was proposed. Four-lead electrocardiogram signals were used to detect abnormal heartbeat conditions, such as premature ventricular complex. 1D CNNs and fully connected layers were utilised in the proposed chip to achieve high-speed, small-area, and high-accuracy arrhythmia detection. The proposed chip was implemented using a 90-nm complementary metal-oxide-semiconductor process and operated at 125 MHz with a core area. The power consumption was at high-speed operation frequency (125 MHz) and at for low-power applications. The detection accuracy was based on the MIT-BIH arrhythmia database. Consequently, the properties of high speed, low power, small area, and high accuracy were established in the proposed accelerator chip. Introduction According to the World Health Organisation, annual mortality from cardiovascular disease is expected to increase from 17.5 million in 2012 to 22.2 million in 2030 [1]. Electrocardiogram (ECG) monitoring is crucial for the prevention and treatment of cardiovascular diseases. ECG monitoring, in the form of a wearable long-term ECG processor, can support the early detection of cardiac arrhythmia [2]. The crucial features of ECG monitoring include low-power use, small area, and high-detection accuracy. In [2], a wearable long-term single-lead ECG processor was presented to support the early detection of cardiac arrhythmia. Only a single-lead ECG was used to perform the detection, with no machine-learning classifier employed; instead, a decision logic based on adaptive thresholds was used for classification. This device, however, is not ideal for use in patients with other heart conditions. A machine-learning-assisted cardiac sensor system-on-chip (SoC) device was also designed for mobile healthcare applications [3]. This design achieved accuracy in detecting arrhythmia and myocardial infarction syndrome with W-level power dissipation for mobile health-care applications. The SoC included low-power biosignal acquisition and a classification system for body sensor networks [4]. With a wavelet transform-based digital signal processing (DSP) circuit and diagnosis control by cardiologists, the accuracy levels of beat detection and ECG classification were close to 99.44 and 97.25%, respectively. An ECG signal processor for the prediction of ventricular arrhythmia using a unique set of ECG features and a naive Bayes classifier was presented in [5]. It utilised a 65-nm CMOS process and its application-specific integrated circuit (ASIC) had an area of and consumed W of power when operated at 10 kHz. The overall classification accuracy reached 86%. A neural network (NN) can be used to detect abnormal heartbeats and classify diseases based on the ECG signals. As shown in Fig. 1, the NN can be trained by both normal and abnormal ECGs, and the test ECG can be classified through a trained NN. In this study, a convolutional NN (CNN) accelerator is proposed to detect the abnormal heartbeat characteristic of premature ventricular complex (PVC) based on four-lead ECG signals. Two parallel CNN filters were designed in the first layer to achieve high-speed computation; two multipliers with corresponding accumulators were designed to execute the fully connected layer, thereby achieving area cost savings. The proposed method was implemented into a single chip by using TSMC 90-nm CMOS processing technology. The results revealed that the proposed core achieved a frequency of 125 MHz with 23.6 K gate counts and power consumption of W at 10 kHz for low-power applications. Furthermore, the detection accuracy was 95.14% based on the MIT-BIH arrhythmia database [6]. Fig 1Open in figure viewerPowerPoint NN for ECG detection Proposed method Cardiac arrhythmia can be monitored through an ECG signal. ECG signals comprise P, QRS complex, T, and U waves. Among these, the QRS complex is the most closely associated with heart rate variability, which is derived from the RR interval of QRS complex contraction and is crucial to the diagnosis of arrhythmia. PVCs are extra abnormal heartbeats that disrupt the regular heart rhythm, sometimes causing the sensation of a skipped beat. PVCs may be diagnosed through the use of a portable ECG for a period of time to capture abnormal heart rhythms. A PVC occurs earlier than a regular heartbeat. Subsequently, the time between the PVC and the next normal beat is longer due to a compensatory pause. Thus, two convolution filters were designed to identify the features of PVCs. Then, the activation function of rectified linear unit (ReLU) and maximum pooling (MaxPooling) was used to produce the ECG data from the convolution filter's output. Finally, the fully connected layer with an activation function of softmax was adopted to calculate the probability of PVC detection. Fig. 2 shows the proposed PVC detection NN. Fig 2Open in figure viewerPowerPoint Proposed NN for PVC detection Fig. 3 shows the architecture of the proposed CNN accelerator. According to the convolution filter, seven shift registers were employed to input the ECG data into the CNN filter. The CNN modules were designed as multipliers and adder trees to enable high-speed computation. The MaxPooling modules were followed by the CNN modules to execute the MaxPooling and ReLU functions. Then, a multiplexer controlled the data to flatten the 2D data, and input to the fully connected computations. Two fully connected modules consisting of two multipliers with corresponding accumulators were designed to obtain a low-cost design. On the basis of the proposed NN, PVCs could be detected accurately. Thus, the proposed CNN accelerator could achieve high-speed, small-area, and high-accuracy abnormal heartbeat detection. Fig 3Open in figure viewerPowerPoint Architecture of proposed CNN accelerator Results and discussion The proposed CNN accelerator was implemented using TSMC 90-nm 1P9M CMOS processing technology. The Synopsys Design Compiler was used to synthesise the register transfer level code and the Cadence Innovus was used for placement and routing. The proposed core was operated at a frequency of 125 MHz and had a power consumption of 4.18 mW. For the low-power application, the proposed chip had a power consumption of W when operated at 10 kHz. The total gate count of the proposed core was 23.6K. The photomicrograph of the proposed core and its characteristics are illustrated in Fig. 4. The MIT-BIH arrhythmia database [6] was applied in the classification test of the proposed CNN accelerator to verify its detection accuracy. The measured results revealed an accuracy of 95.14% for detecting PVCs. Fig 4Open in figure viewerPowerPoint Photomicrograph and characteristics of proposed CNN accelerator Table 1 compares the proposed chip with previous designs. The SoC implementations obtained higher detection accuracy; however, large circuit areas were used in these design types [2-4]. In [2], high classification accuracy was achieved by using simple threshold-based logic. According to the simulation results, this SoC consumed W at a clock frequency of 1 kHz. In [3], a machine-learning processor supported versatile feature extractions and classifications and achieved high accuracy. The measured results demonstrated that the cardiac sensor SoC dissipates W for real-time syndrome detection of ECG-based arrhythmia with 95.8% detection accuracy. In [4], the digital signal processor chip included a monitoring and control service to handle symptom classification. Using this approach, the best detection accuracy was achieved. In [5], the ASIC was implemented using newer technology with a chip size reduction of 45 nm achieved over related chips. Thus, the smallest chip area was achieved. However, this detection method had the lowest accuracy at 86%. The proposed chip adapts CNN to achieve a detection accuracy of 95.14%, which is higher than that of the ASIC design [5] and compares favourably with the performance of the SoC design [3]. In terms of circuit characteristics, the proposed CNN chip has lower power use, a smaller area, and a higher speed design compared with the SoC design. Table 1. Comparison of ECG detection chips DATE'18 [2] JSSC'14 [3] JBHI'15 [4] TVLSI'16 [5] This work technology 180 nm 90 nm 180 nm 45 nm 90 nm type SoC SoC SoC ASIC ASIC results sim. meas. meas. sim. meas. voltage, V 1.0 0.5 1.2 1.0 1.0 area, mm N/A 4.989 2.465 0.112 0.67 method threshold-based machine learning monitoring and control service naïve Bayes CNN detection PVC ventricular arrhythmia/myocardial infarction eight kinds ventricular tachycardia/ventricular fibrillation PVC frequency 1 kHz 25 MHz 120 Hz 10 kHz 125 MHz power, 5.04 48.6 5.97 2.79 3.79@10 kHz accuracy, % 97.02 95.8 97.25 86 95.14 Conclusions To obtain high-detection accuracy, we used a CNN with a four-lead ECG signal to classify PVCs. CNN can accurately extract the features of PVCs. On the basis of the MIT-BIH arrhythmia database verification, the proposed CNN inference accelerator achieved 95.14% detection accuracy. Implemented into a single ASIC chip, the proposed CNN accelerator possesses the characteristics of low power, small area, high accuracy, and high speed. Consequently, the proposed chip is suitable and simple to apply to portable ECG monitors. Acknowledgments This work was supported in part by the Ministry of Science and Technology of Taiwan under project 108-2221-E-182-052 and Chang Gung Memorial Hospital-Linkou under projects CMRPD2G0312, CMRPD2H0301, and CIRPD2F0014. References 1 World Health Organisation: ' Global status report on noncommunicable diseases 2014' ( World Health Organisation, Geneva, Switzerland, 2014) 2Abubakar, S.M., Saadeh, W., Altaf, M.A.B.: ' A wearable long-term single-lead ECG processor for early detection of cardiac arrhythmia'. Design, Automation Test in Europe Conf. and Exhibition (DATE), Dresden, Germany, March 2018, pp. 961– 966 3Hsu, S., Ho, Y., Chang, P. et. al.,: 'A 48.6-to-105.2 W machine learning assisted cardiac sensor SoC for mobile healthcare applications', IEEE J. Solid-State Circuits, 2014, 49, (4), pp. 801– 811 (https://doi/org/10.1109/JSSC.2013.2297406) 4Lee, S., Hong, J., Hsieh, C. et. al.,: 'Low-power wireless ECG acquisition and classification system for body sensor networks', IEEE J. Biomed. Health Inf., 2015, 19, (1), pp. 236– 246 (https://doi/org/10.1109/JBHI.2014.2310354) 5Bayasi, N., Tekeste, T., Saleh, H. et. al.,: 'Low-power ECG-based processor for predicting ventricular arrhythmia', IEEE Trans. Very Large Scale Integr. (VLSI) Syst., 2016, 24, (5), pp. 1962– 1974 (https://doi/org/10.1109/TVLSI.2015.2475119) 6Moody, G.B., Mark, R.G.: 'The impact of the MIT-BIH arrhythmia database', IEEE Eng. Med. Biol. Mag., 2001, 20, (3), pp. 45– 50 (https://doi/org/10.1109/51.932724) Citing Literature Volume56, Issue7March 2020Pages 330-331 FiguresReferencesRelatedInformation
Referência(s)