Depression is a heterogeneous disorder marked by disruptions in cognitive and affective processing. While self-reported measures and clinical interviews remain the diagnostic standard, integrating objective neurophysiological markers could enhance assessment accuracy. This study demonstrates that event-related potentials (ERPs) derived from electroencephalography (EEG) can accurately classify individuals with major depressive disorder (MDD) and predict depression severity.

Participants read multi-sentence scenarios designed to vary in predictability and affective valence, with ERPs time-locked to sentence-final critical words. Features from the Late Frontal Positivity (LFP), N400, and Late Posterior Positivity (LPP) were used to train machine learning classifiers for three tasks: clinical diagnosis (MDD vs. Healthy Controls (HCs)), Beck Depression Inventory-II (BDI)-based depression risk, and Patient Health Questionnaire-9 (PHQ9)-based depression risk.

Our models achieved 80 % accuracy in distinguishing MDD from HC and reliably identified high-risk individuals on both self-reported depression scales. The LPP features were most predictive of clinical diagnosis, whereas N400 and LFP features were more strongly associated with symptom severity. Feature overlap analysis further revealed that distinct neurocognitive processes underlie diagnostic and symptom-based classification, highlighting the potential of these neural markers to capture both categorical and dimensional aspects of depression.

These findings provide compelling evidence that ERPs can serve as objective biomarkers for depression, moving beyond subjective assessments. By leveraging machine learning to analyze neurophysiological responses to linguistic and affective stimuli, this approach lays the foundation for data-driven, personalized psychiatric evaluation—offering a scalable tool for depression diagnosis and severity stratification.

Plain language summary

Major depressive disorder (MDD) is currently diagnosed using self-report measures such as questionnaires asking about symptoms. This is not always reliable if people are unable to define their symptoms. We recruited individuals diagnosed with MDD, and healthy controls to read 2–3 sentence scenarios while we collected data on brain activity. We were able to predict with high accuracy whether individuals were diagnosed with MDD or not, and risk scores on multiple clinical scales. This approach lays the foundation for objective and personalized psychiatric assessments without burdening patients.