Table 3 Feature learning interpretation

Input

2700, 1

–

–

ECG amplitude for one episode

Convolution 1

2700, 1

64

3 \(\times\) 1, stride 1

2698 \(\times\) 64

64 feature map

Convolution 2

2698 \(\times\) 64

64

3 \(\times\) 1, stride 1

2696 \(\times\) 64

64 feature map

Max pooling 1

2696 \(\times\) 64

–

2 \(\times\) 1, stride 2

1348 \(\times\) 64

Feature reduction (1348 nodes for one episode)

Convolution 3

1348 \(\times\) 64

128

3 \(\times\) 1, stride 1

1346 \(\times\) 128

128 feature map

Convolution 4

1346 \(\times\) 128

128

3 \(\times\) 1, stride 1

1344 \(\times\) 128

128 feature map

Max pooling 2

1344 \(\times\) 128

–

2 \(\times\) 1, stride 2

672 \(\times\) 128

Feature reduction (672 nodes for one episode)

Convolution 5

672 \(\times\) 128

256

3 \(\times\) 1, stride 1

670 \(\times\) 256

256 feature map

Convolution 6

670 \(\times\) 256

256

3 \(\times\) 1, stride 1

668 \(\times\) 256

256 feature map

Convolution 7

668 \(\times\) 256

256

3 \(\times\) 1, stride 1

666 \(\times\) 256

256 feature map

Max pooling 3

666 \(\times\) 256

–

2 \(\times\) 1, stride 2

333 \(\times\) 256

Feature reduction (672 nodes for one episode)

Convolution 8

333 \(\times\) 256

512

3 \(\times\) 1, stride 1

331 \(\times\) 512

512 feature map

Convolution 9

331 \(\times\) 512

512

3 \(\times\) 1, stride 1

329 \(\times\) 512

512 feature map

Convolution 10

329 \(\times\) 512

512

3 \(\times\) 1, stride 1

327 \(\times\) 512

512 feature map

Max pooling 4

327 \(\times\) 512

–

2 \(\times\) 1, stride 2

163 \(\times\) 512

Feature reduction (163 nodes for one episode)

Convolution 11

163 \(\times\) 512

512

3 \(\times\) 1, stride 1

161 \(\times\) 512

512 feature map

Convolution 12

161 \(\times\) 512

512

3 \(\times\) 1, stride 1

159 \(\times\) 512

512 feature map

Convolution 13

159 \(\times\) 512

512

3 \(\times\) 1, stride 1

157 \(\times\) 512

512 feature map

Max pooling 5

157 \(\times\) 512

-

2 \(\times\) 1, stride 2

78 \(\times\) 512

Feature reduction (78 nodes for one episode)

Flatten

39,936

–

–

–

Dot product between 78 nodes and 512 feature map

Dense

–

–

–

1000

Weight params

Dense

–

–

–

1000

Weight params

Output

–

–

–

1

Class