Neurophysiological biomarkers for depression classification: Utilizing microstate k-mers and a bag-of-words model.

ABSTRACT

Microstates are analogous to characters in a language, and short fragments consisting of several microstates (k-mers) are analogous to words. We aimed to investigate whether microstate k-mers could be used as neurophysiological biomarkers to differentiate between depressed patients and normal controls. We utilized a bag-of-words model to process microstate sequences, using k-mers with a k range of 1-10 as terms, and the term frequency (TF) with or without inverse-document-frequency (IDF) as features. We performed nested cross-validation on Dataset 1 (27 patients and 26 controls) and Dataset 2 (34 patients and 30 controls) separately and then trained on one dataset and tested on the other. The best area under the curve (AUC) of 81.5% was achieved for the model with L1 regularization using the TF of 4-mers as features in Dataset 1, and the best AUC of 88.9% was achieved for the model with L1 regularization using the TF of 9-mers as features in Dataset 2. When Dataset 1 was used as the training set, the best AUC of predicting Dataset 2 was 74.1% for the model with L2 regularization using the TF-IDF of 9-mers as features, while the best AUC of predicting Dataset 1 was 70.2% for the model with L1 regularization using the TF of 8-mers as features. Our study provided novel insights into the potential of microstate k-mers as neurophysiological biomarkers for individual-level classification of depression. These may facilitate further exploration of microstate sequences using natural language processing techniques.