Assessing the Predictive Utility of Quantitative Electroencephalography Coherence in Adolescent Major Depressive Disorder: A Machine Learning Approach

Abstract

Background:

Improving early recognition and accurate diagnosis of major depressive disorder (MDD) in childhood is a pressing concern. Quantitative electroencephalogram (qEEG) may be an effective, noninvasive diagnostic biomarker for MDD. Prior work by our team demonstrated decreased resting connectivity, as measured by qEEG coherence, in a heterogeneous group of adolescents with MDD compared with age and gender-matched healthy controls (HCs). This study explored qEEG coherence as a predictor of MDD diagnosis in a prospective, longitudinal sample of medication-open-access, adolescents with MDD versus HCs.

Methods:

Twenty-eight adolescents with MDD (Children’s Depression Rating Scale score ≥40) and 27 age and gender-matched HCs (age 14–17, 78% female) received a baseline resting 32-channel EEG. Brain-wide coherence between channel pairs was calculated for the frequency bands (alpha, beta, theta, and delta) and compared between MDD youth and HC. Random forest classifiers were used to predict individual MDD status using baseline qEEG coherence. Models were trained and tested using 10-repeated, 10-fold cross-validation, and performance was evaluated with the area under the receiver operating characteristic curve (AUC-ROC). The contribution of individual predictors was assessed using permutation importance. Model significance was assessed using permutation testing (B = 1000 resamples).

Results:

Random forest models predicted depression status with a trend-level of significance (mean AUC-ROC = 0.65, p = 0.08). Among the most predictive channel pairs, adolescent MDD was characterized by lower coherence in T7-P7 (p < 0.05), Fz-Cz, and Fp2-F8 as well as higher coherence in P4-O2 and Cz-Pz.

Conclusions:

This study provides preliminary evidence that multivariate patterns of qEEG may inform the diagnosis of adolescent MDD. Specific aberrant patterns of coherence within the default mode network and cognitive control network were characteristic of adolescent MDD. Ongoing work will seek to replicate these findings in a larger cohort.