Abstract
Neural networks are increasingly applied
in brain-computer interface research; however, their
decision-making process is often considered difficult to
interpret. This pilot study investigates whether a convolutional neural network can provide not only accu-
rate but also neurophysiologically meaningful classification of electroencephalogram signals recorded during
mirror box training. Two conditions, rest and left hand movement, were analyzed with a focus on interpretability. EEG signals from contralateral electrode C3 were
transformed into time-frequency representations using the Continuous Wavelet Transform, and the resulting spectrograms were used as input to the convolutional
neural network (CNN). To interpret the model’s decisions, Gradient-weighted Class Activation Mapping and other
methods were applied, enabling identification of discriminative time-frequency regions. By relating these pat-
terns to established neuroscientific knowledge, we assess whether the model relies on physiologically meaning-
ful features, providing insight into brain dynamics during mirror-induced motor activity and supporting inter-
pretable BCI applications.
Go back