Anzeige
Erschienen in:
2019 | OriginalPaper | Buchkapitel
Accuracy of Different Machine Learning Type Methodologies for EEG Classification by Diagnosis
verfasst von : Andrius Vytautas Misiukas Misiūnas, Tadas Meškauskas, Rūta Samaitienė
Erschienen in: Numerical Methods and Applications
Aktivieren Sie unsere intelligente Suche, um passende Fachinhalte oder Patente zu finden.
Wählen Sie Textabschnitte aus um mit Künstlicher Intelligenz passenden Patente zu finden. powered by
Markieren Sie Textabschnitte, um KI-gestützt weitere passende Inhalte zu finden. powered by
Abstract
Electroencephalogram (EEG) classification accuracy of different automatic algorithms (including their setup) is discussed. Two patient groups, characterized by visually similar (to neurologists) EEG rolandic spikes, are under classification. The first group consists of patients with benign focal childhood epilepsy. Patients with structural focal epilepsy define the second group. We analyzed 94 EEGs (with known diagnosis) obtained from Children’s Hospital, Affiliate of Vilnius University Hospital Santaros Klinikos.
The EEGs are preprocessed by applying these steps: (i) spike detection; (ii) extraction of spike parameters. After preprocessing of EEGs we gather parameters of detected spikes into lists of equal length \(N_{spikes}\).
The classification algorithms are trained employing one set of patients (containing patients from both groups) and tested on another non-overlapping set of patients (also from both groups). This prevents artificial accuracy inflation due to overfitting.
We compared eight machine learning type classifiers: (1) random forest, (2) decision tree, (3) extremely randomized tree, (4) adaptive boosting (AdaBoost), (5) artificial neural network (ANN), (6) supported vector machine (SVM), (7) linear discriminant analysis (LDA), (8) logistic regression. To estimate quality of classifiers we discuss a set of metrics. The results are following: (I) as expected, for all examined algorithms, the accuracy tends to grow (when \(N_{spikes}\) increases), saturating at some asymptotic value; (II) ANN has prevailed as best classifier.
Impact of: (a) different training strategies and (b) spike detection errors on classification accuracy is also discussed.
Novelty and originality of this study comes not only from classifying different types of epilepsy, but also from employed computational methodology (involving parameters of EEG spikes and machine learning type classifier), as well as comparing different methodologies of such type, based on their accuracy and other classifier metrics.