Table 1.

Architecture of a 3D convolutional neural network used to distinguish obstructive apnea from central apnea.

Layer	Number of filters, n	Size/stride	Activation function	Output size
Input	N/Aa	N/A	N/A	480×640×20×2
Average pool	N/A	25×25×1/20×20×1	N/A	23×31×20×2
Convolutional	8	2×2×1/1×1×1	Linear	22×30×20×8
Dropout	N/A	N/A	N/A	22×30×20×8
Convolutional	16	3×3×5/1×1×1	N/A	20×28×16×16
Max pool	N/A	8×8× /2×2×1	N/A	7×11×16×16
Batch normalization	N/A	N/A	Leaky Relub	7×11×16×16
Convolutional	64	2×2×2/1×1×1	N/A	6×10×15×64
Batch normalization	N/A	N/A	Leaky Relu	6×10×15×64
Convolutional	32	4×4×1/1×1×1	N/A	3×7×15×32
Batch Normalization	N/A	N/A	Relu	3×7×15×32
Dropout	N/A	N/A	N/A	3×7×15×32
Convolutional	16	2×2× /1×1×1	N/A	2×6×15×16
Batch normalization	N/A	N/A	Relu	2×6×15×16
Flatten	N/A	N/A	N/A	2880
Fully connected	16	2880×16	N/A	16
Fully connected	4	16×4	N/A	4
Output layer	N/A	4×1	Sigmoid	1

^aN/A: not applicable.

^bReLu: rectified linear unit.