Arousal Detection

Windowing

Window size is 30 seconds, Windows have NOT Overlap in training

Windowing Process

• Move Window (15 seconds with overlap, 30 without overlap)
• Find the most labels as Y
• Save in file( x, x_prev(-1), x_next(-1))

Input

The input of the Neural network is 1400 samples that is 28 seconds(Data downsampled to 50Hz) Windows size is 30 seconds(1500 sample).

How to Get 1400 samples from 30 seconds

Cropping If the length of a window is more than 1400 samples then choose 1400 consecutive samples from it randomly.

Batch-Size

The batch size is a number of samples processed before the model is updated Batch-size is 32 Chunk Batch-size/2 is positive windows(most labeled as 1) Batch-size/2 is negative windows(most labeled as 0)

Chunk

Chunk contains X_Prev, X, X_Next after pre-processing.

How to choose Batch-size/2

Choose batch-size/2 of users randomly, choose on of the positive or negative windows from each users.

Other Pre-Processings

Arch

CNN for channel except SaO2: CNN for SaO2:

Convolution layers

Example

Max-Pooling

1D max-pooling Pool-size = 2 Stride = 1

PReLU

As activation

Drop-Out

Rate = 0.33 The Dropout layer randomly sets input units to 0 with a frequency of rate at each step during training time, which helps prevent overfitting. Inputs not set to 0 are scaled up by 1/(1 - rate) such that the sum over all inputs is unchanged.

Flatten layer

To remove all of the dimensions except for one

Fully Connected layer

There are 3 Dense layers that after each one there is a PReLU layers and Dropout layer. All channels except SaO2 have [256,128,128] units and SaO2 has [512,256,64,64] units

Adam Optimizer

Adam optimizer with learning rate 0.00005

Model Result

Sequential