| dc.description.abstract | Electroencephalogram (EEG) in combination with machine learning (ML) techniques is becoming an increasingly popular method in medicine for clinical disorders prediction. This study applies ML techniques to the ePod dataset developed by the ePODIUM project, for the prediction of developmental dyslexia in infants. The dataset contains EEG recordings of 129 infants, existing out of two groups with dyslexic- and non-dyslexic parents, obtained from an experiment with auditory stimuli for eliciting a Mismatch Negativity (MMN). Four different approaches for feature selection are used to see the differences in performance on different ML algorithms. The baseline approach uses the MMN of all EEG channels. The second approach includes EEG channels reported in the literature as the most informative. The t-test approach uses significance testing, verified using a t-test on the ePod data, and resulted in a selection of significantly different channels between the two groups. The final approach uses the channels from the ePod data that show the highest connectivity with other channels. The algorithms used on the different feature input approaches are support vector machine, logistic regression, decision tree, multilayer perceptron, and convolutional neural network. The convolutional neural network showed the highest performance in combination with the features of the t-test approach with an
accuracy of 73%. However, this result is not significant (p=0.447) because of high variation in model performance. The connectivity approach performs also well based on average accuracy with the convolutional neural network. The traditional machine learning algorithms support vector machines and logistic regression can learn from the t-test and connectivity with moderate accuracy of 60%. The results show that data-driven selected features, using significance testing and connectivity, are promising in predicting developmental dyslexia in infants in combination with deep learning and traditional machine learning models. | |
