SR-ScatNet Algorithm for On-device ECG Time Series Anomaly Detection

IEEE SoutheastCon 2021 – Engineers Connecting the World
Date of Conference: 10-13 March 2021
Conference Location: Atlanta, GA, USA

Transcript

1. IEEE Region 3 SoutheastCon 2021 1 SR-ScatNet Algorithm for On-device ECG Time Series Anomaly Detection Hsin-Yu Feng, Po-Ying Chen, Janpu Hou March 12, 2021 March 12, 2021
2. 2 Abstract: Anomaly detection of real-time ECG time series is of particular interest for early detection of cardiovascular disease for aging population. To use convolutional neural networks (CNN) for any on- device training or inference, you need GPU-accelerated hardware which will not only increase the hardware cost but also consume higher battery power. We proposed a SR-ScatNet algorithm for on-device application such as smart cloth with ECG monitoring sensors. Two improvements were made. First on spectral residual, we use Fourier Transform of autocorrelation of ECG signals instead of original time series to increase the sensitivity. Second on feature extraction, we use shallow wavelet scattering network (ScatNet) instead of deep CNN network so the on-device training can be performed on a simple Arm Cortex-A53 processor without any GPU-accelerator. These improvements are made to create a compact machine learning model according to the nature of different waves constituting the ECG signals. To verify the proposed method, we use the MIT-BIH Arrhythmia Database. The spectral residual of autocorrelation ECG signals can detect the abnormal ECG signals with over 98% accuracy. The wavelet scattering network can further classify the type of abnormality with over 90% accuracy. We believe the design of ECG monitoring smart cloth can benefit from such SR-ScatNet algorithm.
3. Outline ➢Introduction • Chronic Diseases Home Health Monitoring • ECG smart cloth ➢Method • The power density spectral residual • The simplified wavelet scattering network • Dataset ➢Results and Discussion • Abnormality detection Evaluation • Classification Evaluation ➢Conclusion 3
4. Remote Patient Monitoring for Chronic Diseases Control Home Health Hub 4 Body Fat Scale BPM Oximeter Glucose meter Spirometer Thermometer ECG Emergency Hospital Clinician Home Health Devices intervention Edge Cloud (Sense) (Infer, Act)
5. The need of Intelligent at the Edge ➢Data security, data privacy ➢Area with limited or low cloud access ➢Smartly minimize transmission bandwidth • Critical vs non-critical data • Informational vs actionable • Routine vs abnormal trends ➢Edge device resource usage optimization • Storage-memory-compute balance Advantage to have On-device machine learning model 5
6. ECG 6 Aortic Pressure Artery Pressure Ventricular pressure Ventricular volume ECG
7. Atrial Fibrillation (AFiB) Premature Ventricular Contractions (PVCs) Normal Sinus Rhythm (NSR)
8. ADS1115 AD8232 ARM Cortex-A53 ECG abnormality Warning System Inside view Front view Smart Cloth ECG Monitoring System
9. Outline ➢Introduction ➢Method • The power density spectral residual • The simplified wavelet scattering network • Dataset ➢Results and Discussion ➢Conclusion 9
10. FEATURE EXTRACTION ABNORMAL DETECTION Scattering Coefficients Power Density Spectral Residual Normal ECG sensor yes no Spectral Residual ScatNet Support Vector Machine CLASSIFICATION SR-ScatNet Framework for ECG Anomaly Detection
11. Spectral Density Estimation ➢PSD: distribution of power over the entire frequency range. ➢Step 1: Find ECG signals’ auto-correlation function ➢Step 2: Take Fourier Transform ➢Step 3: Compute Spectral Residual Power Spectral Density Features 11 Incoming ECG spectrum Average Normal ECG spectrum ECG Spectral Residual = - Note: Convolutional Neural Networks (CNN) is using correlation to find similarity
12. Original ECG signals Normal Abnormal Abnormal Normal PVC AFiB Power Spectral Density Residual
13. ScatNet: similar architecture simpler computation 13 Input ECG Signal Conv ReLu Pooling Conv - 1 Conv ReLu Pooling Conv - 2 Conv ReLu Pooling Conv - 3 FC ReLu FC-1 Softmax Feature Extraction Classifier Input ECG Signal Conv Modulus Low Pass Filter Scat - 1 Conv Modulus Low Pass Filter Scat - 2 Conv Modulus Low Pass Filter Scat - 3 SVM Feature Extraction Classifier Convolutional Neural Network Scattering Convolution Network Pre-defined Wavelet Filter
14. ScatNet Representation of ECG Signal 14
15. ScatNet Representation of ECG Signal PVC AFiB Normal
16. ➢Normal • NOR, PAB ➢Low Risk • LBBB, RBBB, VEB ➢Intermediate Risk • PVC, PAC ➢High Risk • VFW, AFiB ECG Dataset: MIT-BIH Arrhythmia Database Static State – NOR, PAB Dynamic State – During Exercise Steady State Over Time – PVC, PAC
17. Outline ➢Introduction ➢Method ➢Results and Discussion • Abnormality detection Evaluation • Classification Evaluation ➢Conclusion 17
18. Abnormality detection Evaluation
19. Classification Evaluation 19 1396 (98%) 0 0 0 0 0 28 (2%) 98.4% 1.96% 393 (28%) 1010 (72%) 0 0 0 0 0 72.1% 27.9% 0 0 1582 (98%) 0 0 0 32 (2%) 100% 0.0% 0 0 0 94 (100%) 0 0 0 100% 0.0% 0 0 0 0 361 (71%) 46 (9%) 101 (20%) 70.56% 29.44% 0 0 0 0 28 (2%) 1112 (78%) 256 (18%) 78.09% 21.91% 0 0 200 (2%) 0 400 (4%) 100 (1%) 9303 (93%) 93.00% 7.00% 90.25% 9.75% AFiB AFL LBBB WPW PAC PAV NSR Predicted Label AFiB AFL LBBB WPW PAC PVC NSR Low Risk Medium Risk High Risk True Label
20. 2799 (99%) 0 28 (1%) 99% (1%) 0 1676 (98%) 32 (2%) 98% (2%) 0 200 (1.8%) 10717 (98.2%) 98.2% (1.8%) 98.3% (1.7%) Low Risk Medium Risk High Risk True Label Low Risk High Risk Medium Risk Predicted Label Accuracy = 98.3%, Misclass = 1.7% Abnormality detection Evaluation
21. Conclusion 21 ECG Smart Cloth Gateway Device ECG Smart Cloth Gateway Device Intelligent Cloud Intelligent Edge Hardware: CPU+GPU Software: TensorFlow ConvNet Hardware: MCU Software: FFT SR-ScatNet Sense Infer, Act Sense, Infer, Act Learn Intelligent Cloud
22. Conclusion ➢We demonstrated a two-stage algorithm for ECG wearable devices for abnormality detection. ➢By combining the spectral residual and wavelet scattering network, a Fast Fourier Transform (FFT) based SR-ScatNet algorithm required much less computation and can be done on smart cloth ECG sensor platform. ➢Such on-device application can mitigate network limitations, reduce energy consumption, increase security, and improves data privacy due to the training of learning model is done right at device where the data is. ➢The detection itself can be done on device without the need to transfer all the ECG data to the cloud. 22
23. 23 Who Has Cited This Publication?