Figure 2
From: A robust deep learning detector for sleep spindles and K-complexes: towards population norms
SEED’s event detection process and architecture. (a) Raw EEG signals are preprocessed and segmented in windows of \({T}_{w}\)=20 s samples. Each window is fed to the deep neural network underlying SEED, which estimates each sample’s probability of being part of an event (SS or KC). To avoid border effects, signal segments of \({T}_{B}\) samples are concatenated to both input window’s borders, which are dropped afterward. The raw probabilities are adjusted and thresholded using a pair of thresholds, one for detection and another for duration estimation. The resulting events can be post-processed using expert knowledge. Finally, events outside the valid annotation mask (e.g., N2 stages) are discarded. (b) High-level description of SEED's neural network architecture. (c) Definition of Convolutional Multi-dilated Block (Conv MDB) k, F. It has the same number of parameters as a sequence of two Conv k, F layers, but with a larger receptive field.
