ECG Beat Classification on Edge Device

2020 IEEE International Conference on Consumer Electronics (ICCE)
Date of Conference: 04-06 January 2020
Conference Location: Las Vegas, NV, USA

Transcript

1. IEEE Consumer Electronics Society ICCE 2020 January 6, 2020 Session 3.8 CHS (1), Room G, 17:20 ECG Beat Classification on Edge Device Presenter: Janpu Hou
2. IEEE Consumer Electronics Society. IEEE International Conference Consumer Electronic ICCE 2020 2 Introduction – ECG Signal Representation and Classification – Smart Clothing Deep Learning CNN Architecture – Wavelet Scattering Transform – Subspace Approximation with Augmented Kernels (Saak Transform) Results and Discussion Outlines:
3. IEEE Consumer Electronics Society. IEEE International Conference Consumer Electronic ICCE 2020 3
4. IEEE Consumer Electronics Society. IEEE International Conference Consumer Electronic ICCE 2020 4 Time series analysis: – CNN architecture with Saak Transform Simple On-device Training: – No deep learning frame work such as PyTorch, TensorFlow required Low Power Inference: – No accelerator such as Google Coral, Intel Movidus VPU required ECG abnormal warning system with Deep Learning Model
5. IEEE Consumer Electronics Society. IEEE International Conference Consumer Electronic ICCE 2020 5 Normal – NOR, PAB Low Risk – LBBB, RBBB, VEB Intermediate Risk – PVC, PAC High Risk – VFW ECG Tracking Smart Clothing Static State – NOR, PAB Dynamic State – During Exercise Steady State Over Time – PVC, PAC
6. IEEE Consumer Electronics Society. IEEE International Conference Consumer Electronic ICCE 2020 6 Heart's Electrical System Smart Cloth ECG Monitoring System ECG Abnormal Warning System
7. IEEE Consumer Electronics Society. IEEE International Conference Consumer Electronic ICCE 2020 7 Personalized Arrhythmia Monitoring Platform ADS1115 AD8232 ARM Cortex-A53 Prof. Shuenn-Yuh Lee, NCKU, 裕晶醫學科技 YuTech Co., Ltd.
8. IEEE Consumer Electronics Society. IEEE International Conference Consumer Electronic ICCE 2020 8 Introduction – ECG Signal Analysis – Smart Clothing Deep Learning CNN Architecture – Wavelet Scattering Transform – Subspace Approximation with Augmented Kernels (Saak Transform) Results and Discussion Outlines:
9. IEEE Consumer Electronics Society. IEEE International Conference Consumer Electronic ICCE 2020 9 Layer-to-layer Transform • A linear operator W with • A pointwise Non-linearity r A Standard CNN Architecture Filter Weights optimized by backpropagation with respect to a given task Bruna and Mallat
10. IEEE Consumer Electronics Society. IEEE International Conference Consumer Electronic ICCE 2020 10 A Scattering Transform Architecture • Wavelet Transform: Convolution • Averaging: Max pool • Modulus: Non-linear Activation Filter Weights Mathematically Pre-defined Bruna and Mallat
11. IEEE Consumer Electronics Society. IEEE International Conference Consumer Electronic ICCE 2020 11 Scattering networks are a class of designed CNN with fixed filter banks. Sparse (one-stage) – Find Invariants Stable (multi-stages) – Windowed scattering transform for any wavelet stability to deformations Classify Scattering Invariants Key idea: Find accurate representations with few parameters Great! But require Pytorch or TensorFlow and GPU
12. IEEE Consumer Electronics Society. IEEE International Conference Consumer Electronic ICCE 2020 12 A Standard CNN Operation
13. IEEE Consumer Electronics Society. IEEE International Conference Consumer Electronic ICCE 2020 13 augmented kernels One stage Saak Transform
14. IEEE Consumer Electronics Society. IEEE International Conference Consumer Electronic ICCE 2020 14 A Saak Transform Architecture 64x64x3 32x32x8 16x16x12 8x8x16 4x4x24 2x2x32 • Saak Transform: Convolution • Remove low variance: Max pool • ReLU + Augmented Kernels : Non- linear Activation Data-driven Filter Weights
15. IEEE Consumer Electronics Society. IEEE International Conference Consumer Electronic ICCE 2020 15 Outlines: Introduction – ECG Signal Analysis – Smart Clothing Methodology – Wavelet Scattering Transform – Successive Subspace Learning Results and Discussion
16. IEEE Consumer Electronics Society. IEEE International Conference Consumer Electronic ICCE 2020 16 ECG Signal Type – Annotation Clusters 1. Normal 75017 2. Left Bundle Brunch Block 8072 3. Right Bundle Brunch Block 7255 4. Premature Atrial Contraction 2546 5. Premature Ventricular Contraction 7129 6. Paced Beat 7024 7. Ventricular Flutter 472 8. Ventricular Escape Beat 106
17. IEEE Consumer Electronics Society. IEEE International Conference Consumer Electronic ICCE 2020 17
18. IEEE Consumer Electronics Society. IEEE International Conference Consumer Electronic ICCE 2020 18
19. IEEE Consumer Electronics Society. IEEE International Conference Consumer Electronic ICCE 2020 19 Sensitivity (true positive rate) = 98.3 % Specificity (true negative rate) = 99.1% Abnormal Detection:
20. IEEE Consumer Electronics Society. IEEE International Conference Consumer Electronic ICCE 2020 20 Time series analysis – CNN architecture solved with a PCA-based subspace approximation (Saak Transform) Simple On-device training – Anywhere – Anytime Low-power consumption – No AI accelerator Summary:
21. IEEE Consumer Electronics Society. IEEE International Conference Consumer Electronic ICCE 2020 21 Apply to other Bioelectrical Signals – EGG, EEG, EMG Apply Topological data analysis – Transform ECG signal into a graph Distributed-Feedback model – Update the model not the personal data Future Work