A Machine Learning-Based Analysis of EEG Biomarkers for Predicting Antidepressant Treatment Response in Major Depressive Disorder

Major depressive disorder (MDD) is often treated via trial-and-error, as few objective biomarkers exist to predict which patients will respond to antidepressants. This study investigates whether machine learning (ML) models can predict 12-week antidepressant treatment response by analyzing pre-treatment and 1-week post-treatment clinical data and various electroencephalographic (EEG) metrics.

The researchers followed 51 adults with MDD over a 12-week pharmacotherapy trial. Participants were classified as responders (≥50% reduction in depression scores) or non-responders. The researchers developed ML models (primarily Random Forests) using three main data sources: (a) scalp-level EEG power, (b) eLORETA source-localized brain activity, and (c) demographic and clinical symptom data. The analysis progressed through feature selection (to remove noise), feature mapping (to reduce dimensionality), and final classification, comparing individual models against a combined model that included all relevant predictive features.

The study demonstrated that EEG features and clinical data provide significant predictive value when analyzed together. The most predictive individual features included baseline frontopolar theta power and specific alpha band activity. When all relevant features were combined into a single model, the accuracy for predicting treatment response reached 88%. Even after narrowing the input down to the 12 most important features, the model maintained a classification accuracy of 78%. Notably, the model showed high specificity, suggesting it is particularly effective at identifying individuals who may not respond to initial treatment, thus potentially avoiding ineffective trials.

This research provides a framework for transitioning from trial-and-error antidepressant prescribing to a more personalized, biomarker-guided approach. By demonstrating that relatively simple EEG markers and specific clinical symptoms like concentration difficulty can be processed by ML to predict outcomes early in the treatment course, this approach holds promise for improving the speed and efficacy of psychiatric care.