Title Page
Contents
국문 요약 11
ABSTRACT 13
1. INTRODUCTION 15
2. BACKGROUND 22
2.1. Recurrent Neural Network 22
2.2. Long Short-Term Memory 25
2.3. Bidirectional Long Short-Term Memory 28
2.4. Convolutional Neural Network 29
2.5. Dilated Causal Convolution 34
2.6. Attention Mechanism 37
2.6.1. Dot-Product Attention 37
2.6.2. Self-Attention 42
3. RELATED WORK 46
3.1. Power Demand Prediction 46
3.1.1. Material 51
3.1.2. Methodology 55
3.1.3. Experimental Results 59
3.2. Atrial Fibrillation Classification 64
4. Paroxysmal Atrial Fibrillation Classification 68
4.1. Material 68
4.2. Methodology 73
4.2.1. R-peak Detection 73
4.2.2. TS-ECG Details 79
4.3. Experimental Results 88
4.3.1. Experimental Environment 89
4.3.2. Result of R-peak Detection 90
4.3.3. Classification Experimental Results 96
5. CONCLUSION 111
6. REFERENCE 112
Table 1. 〈Key, Context[₁,c]〉 pair notation[이미지참조] 51
Table 2. Datasets utilized for the training procedures 54
Table 3. Range of hyperparameters 57
Table 4. Experimental results to determine the number of layers l with d₁ 60
Table 5. Error metric results for with holiday datasets 61
Table 6. Error metric results for without holiday datasets 62
Table 7. Weekday/weekend dataset results without context. l-LSTM is an ensemble model. 63
Table 8. The patient information from the Inha University Hospital dataset 69
Table 9. Patient information from another hospital used as an external test dataset 72
Table 10. Classification results of the dataset A. Support is the number of occurrences of each class when it is true 98
Table 11. Experimental results determining the optimal value of m for dataset A 107
Table 12. Evaluation of the performance of each framework for TS-ECG on dataset A. TS-ECGT is a model trained exclusively on total length ECG data. TS-ECGS is a model trained exclusively on P-wave segments data.[이미지참조] 107
Table 13. TS-ECG performance by dilation rate 108
Table 14. Results of applying dilation rates with conventional convolution operations on TS-ECG 108
Table 15. Classification results on external test datasets B and C 110
Figure 1. Multi-layer perceptron 16
Figure 2. Lead II of ECG representations: (a) Normal sinus rhythm (NSR) in healthy individuals (b) NSR in individuals with paroxysmal atrial fibrillation (PAF) 20
Figure 3. Structure of RNN 22
Figure 4. Graph of the hyperbolic tangent function and its derivative 24
Figure 5. Structure of LSTM cell 25
Figure 6. Structure of LSTM network 26
Figure 7. Structure of Bi-LSTM network 28
Figure 8. Operation of a convolution layer for a single channel: (a) Convolution operation result when stride=1 (b) Convolution operation result when stride=2 30
Figure 9. Padding operation in single channel convolution layer: (a) In the case without padding (b) In the case with padding 31
Figure 10. Example of applying max-pooling with stride=2 32
Figure 11. Representation of a series of causal convolutions 35
Figure 12. Example of applying dilation rate to a convolution filter 35
Figure 13. Representation of a series of dilated causal convolutions 36
Figure 14. Structure of sequence to sequence model 37
Figure 15. Attention score through dot-product operation 38
Figure 16. Attention distribution obtained by applying the softmax function to the attention scores 39
Figure 17. Attention value computed as the sum of the products of attention weights and hidden states for each word 40
Figure 18. Compute vt by concatenating the attention value with the decoder's hidden state at time-step t[이미지참조] 41
Figure 19. Conceptual diagram of the attention function 42
Figure 20. Self-attention operation process: (a) Process of deriving the query, key, and value vectors (b) Process of constructing the attention score matrix (c) Process... 43
Figure 21. Overview of S2S model structure for power demand prediction 48
Figure 22. Structure of CLDNN model 49
Figure 23. CNN-LSTM hybrid model structure 50
Figure 24. Sliding window and 〈Key, Context[₁,c]〉 pair[이미지참조] 52
Figure 25. Analysis of power demand: (a) yearly mean power usage across seasons (b) weekly average power demand distinguished by weekdays and weekends 53
Figure 26. LSTM network composed of l layers 56
Figure 27. Architecture of our proposed hybrid model. The output features from the c LSTM layers are fed into the CNN 56
Figure 28. Overview model structure for extracting and training with the P-wave 65
Figure 29. ECG beat 66
Figure 30. ECG data of 8-lead 68
Figure 31. Disease statistics for Inha University Hospital dataset 70
Figure 32. ECG data of 12-lead created using Einthoven's triangle 70
Figure 33. Disease statistics for all hospital test datasets 72
Figure 34. First framework: (a) Residual connection with dilated causal convolution (b) Composition of one residual block. We use k=3 and d is a multiple of 2 (c)... 81
Figure 35. Effect of each convolution operation 83
Figure 36. Second framework: (a) Composition of Bi-LSTM module (b) Composition of one Bi-LSTM block (c) Feature extraction from sub-length ECG Waveforms 85
Figure 37. The overall structure of the TS-ECG model 86
Figure 38. The entire progression of the proposed method 87
Figure 39. R-peak detection using the proposed Algorithm 4 91
Figure 40. R-peak detection using the Pan and Tompkins algorithm 92
Figure 41. R-peak detection using the Engelse and Zeelenberg algorithm 93
Figure 42. R-peak detection using the stationary wavelet transform 94
Figure 43. Extracting P-wave segments with the proposed Algorithm 6 95
Figure 44. Grad-cam results of the first framework 104
Algorithm 1. Calculating distance function algorithm 73
Algorithm 2. Select candidate function algorithm 74
Algorithm 3. Check RR interval function algorithm 74
Algorithm 4. Extracting P-wave segments function algorithm 75
Algorithm 5. Cosine similarity check function algorithm 76
Algorithm 6. Make sub waves function algorithm 77
Equation 2-1. [제목없음] 23
Equation 2-2. [제목없음] 23
Equation 2-3. [제목없음] 26
Equation 2-4. [제목없음] 26
Equation 2-5. [제목없음] 26
Equation 2-6. [제목없음] 26
Equation 2-7. [제목없음] 26
Equation 2-8. [제목없음] 26
Equation 2-9. [제목없음] 26
Equation 2-10. [제목없음] 29
Equation 2-11. [제목없음] 31
Equation 2-12. [제목없음] 32
Equation 2-13. [제목없음] 33
Equation 2-14. [제목없음] 39
Equation 2-15. [제목없음] 40
Equation 2-16. [제목없음] 40
Equation 2-17. [제목없음] 44
Equation 3-1. [제목없음] 59
Equation 3-2. [제목없음] 59
Equation 4-1. [제목없음] 69
Equation 4-2. [제목없음] 69
Equation 4-3. [제목없음] 69
Equation 4-4. [제목없음] 69
Equation 4-5. [제목없음] 82
Equation 4-6. [제목없음] 88
Equation 4-7. [제목없음] 88
Equation 4-8. [제목없음] 88
Equation 4-9. [제목없음] 88