Máté Baradits1, Kakuszi Brigitta1, Sára Bálint1, István Bitter1, Pál Czobor1
1 Semmelweis University Department of Psychiatry and Psychotherapy, Budapest
Introduction: Diagnosis of psychiatry disorders are biased by the subjective evaluation of the psychiatrist, based on the observed behaviors of the subject. Machine learning techniques offer new opportunities for an objective classification of psychiatric disorders based on neuroimaging data such EEG or fMRI. Here we use microstate segmentation - proposed by Lehmann et al. - which is a method that considers EEG recordings as spatial topographies and clusters them to microstates. These microstates can be described with their average duration, occurrence per second (frequency), total coverage of time and probabilities of transition from one to another.
Aims: The goals of our study were to (1) use microstate segmentation to extract features from resting state EEG activity, and (2) to use these features in classification models.
Methods: Data were acquired with a high-density, 256 channel Biosemi-EEG system, in rest with closed eyes. We investigated 70 patients with schizophrenia and 75 healthy controls. For microstate segmentation K-mean clustering algorithm was used. We used linear discriminant analysis (LDA) and Support Vector Machine (SVM) to investigate how well the microstate measures can distinguish the two groups.
Results: In patients with schizophrenia, microstate class ‘A’ duration was shorter, occurred less frequently and covered a lower percentage of time, while microstate class ‘B’ and ‘D’ had longer duration, occurred more frequently and covered more percentage of time. LDA was able to distinguish the two groups with approximately 76% mean accuracy (cross-validated) for both groups, while SVM reached approximately 81% mean accuracy.
Conclusion: There is a difference in patients with schizophrenia and healthy controls in microstate measures. Features derived through microstate segmentation were able to distinguish schizophrenia patients from healthy controls, where SVM’s accuracy was superior to LDA’s. Our findings indicate that machine learning techniques may provide useful tools to supplement psychiatric diagnosis.
Doctoral School of Mental Health Sciences
Supervisor: Pál Czobor
E-mail address: firstname.lastname@example.org