Machine learning with electroencephalography features for precise diagnosis of depression subtypes

Abstract: Depression is a common psychiatric disorder, which causes significant patient distress. Bipolar disorder is characterized by mood fluctuations between depression and mania. Unipolar and bipolar depression can be easily confused because of similar symptom profiles, but their adequate treatment plans are different. Therefore, a precise data-driven diagnosis is essential for successful treatment. In order to aid diagnosis, research applied machine learning to brain imaging data, in particular to electroencephalography (EEG), with accuracies reaching 99.5% (unipolar vs. healthy) or 85% (bipolar vs. healthy). However, these results arise from small training sets, without validation on independent data, and thus have a high risk of inflated accuracies due to data over-fitting. We propose to use a bigger corpus of realistic clinical data for training and testing and improve classification with microstates features, which can assess the function of large-scale brain networks.