Machine learning-based source-level EEG functional network biomarkers of antidepressant treatment response in major depressive disorder.

Baseline EEG network features distinguished later antidepressant responders from nonresponders within this research sample. The model has not yet been validated as a clinical test for selecting medication.

Song YW, Lee HA, Kim S, Kim JS.
Progress in Neuro-Psychopharmacology & Biological Psychiatry · Publication date: 2026-10-02
Added to site: 2026-10-03
DOI: 10.1016/j.pnpbp.2026.111957

Publication date verified in PubMed. This source date is separate from the site's addition date.

Modalities

EEG, Source-level EEG, Machine learning

Clinical focus

Major depressive disorder, Treatment-response biomarkers, Functional connectivity

Population

Patients with major depressive disorder and a healthy comparison group; age and treatment-regimen details were not exposed in the checked abstract

Sample

47 patients with major depressive disorder and 41 healthy controls; 20 patients were classified as responders and 27 as nonresponders

Intervention / method

Baseline source-level EEG networks were characterized with graph measures, then nodal features were used in nested cross-validation to classify response after 12 weeks

Comparator

Antidepressant responders, defined by at least 50% Hamilton Depression Rating Scale reduction, versus nonresponders; healthy controls informed selected comparisons

Clinician summary

A support-vector machine using nodal network features reached 82.98% accuracy, 85.19% sensitivity and 80.00% specificity in nested cross-validation. High-beta network measures differed between responders and nonresponders, and responder theta-band clustering differed from healthy controls.

How to interpret the findings

Only 47 patients contributed outcome labels, and the model was not tested in an independent external cohort. The abstract does not expose medication standardization, preprocessing choices, feature-selection stability or calibration. Internal cross-validation performance does not establish prospective treatment-selection utility.

Equipment and disclosures

Resting-state EEG and source-level network analysis were used. The checked abstract does not identify the amplifier, channel count, montage, software or commercial system.

The authors declared no known competing interests. Crossref identifies support from the National Research Foundation of Korea and Korea Ministry of Health and Welfare.

Related evidence

Compare findings from the same clinical area, including mixed results.

• Intrinsic functional architecture predicts remote motor cortical excitability changes induced by intermittent theta-burst stimulation · Positive with limitations
• Frontal midline theta reduction during attentional reconfiguration links local cortical dynamics to network-level connectivity changes · Positive with limitations
• Brain state-contingent stimulation to probe task-dependent functional networks. · Positive with limitations
• Neuromodulation effects of HF-rTMS and rPMS on stroke patients: an fNIRS study. · Positive with limitations
• Targeting the dopaminergic midbrain with precision 7-Tesla biofeedback training in depression: A proof-of-principle randomized controlled trial. · Positive with limitations
• EEG Microstate Alterations in Anhedonia Subtypes of Major Depressive Disorder and Their Modulation by Intermittent Theta-Burst Stimulation. · Positive with limitations
• Temporal-Targeted Accelerated 40 Hz tACS Enhances Cognitive Function and Network Integration in Alzheimer's Disease: A Randomized Clinical Trial. · Positive with limitations
• Cerebellar Theta-Burst Stimulation Induces Distinct Functional Connectivity Patterns in Post-Stroke Aphasia: A Randomized Controlled Trial. · Mixed
• Infra-low-frequency neurofeedback alters EEG network efficiency: exploratory evidence from healthy volunteers. · Positive with limitations
• Escaping the Negative Attentional Bias in Depression with Real-Time Neurofeedback · Positive with limitations