Real-Time EEG Neurofeedback Using a Subject-Independent Classifier in Women With Depressive Symptoms

Classifier accuracy was the percentage of classifier outputs during regulation blocks that matched the instructed mental state. Each run contained 160 classifier outputs during regulation periods. Run-level accuracy was averaged across four runs to obtain a session-level value, and session-level accuracy was averaged across five sessions for each participant. Overall accuracy was compared with the theoretical 50% chance level within each arm and between arms.

For each participant, the percentage of classifier outputs matching the instructed mental state was averaged across four runs separately for Sessions 1, 2, 3, 4, and 5. Session-specific accuracy was compared with the theoretical 50% chance level within each arm.

Early-training accuracy was the mean of Sessions 1 and 2. Late-training accuracy was the mean of Sessions 4 and 5. Each phase was compared with the theoretical 50% chance level, early-to-late change was assessed within each arm, and phase-specific values were compared between arms.

The Hamilton Depression Rating Scale was administered by trained clinical psychologists masked to allocation. Lower scores indicate fewer clinician-rated depressive symptoms. Change was calculated as baseline score minus the later score; positive values indicate symptom improvement. Analyses used available paired observations without imputation.

The Beck Depression Inventory-II is a self-reported measure of depressive symptom severity. Lower scores indicate fewer symptoms. Change was calculated as baseline score minus the later score; positive values indicate symptom improvement. Analyses used available paired observations without imputation.

Correlations were calculated separately by arm between classifier performance and clinician-rated depressive symptom change measured with the Hamilton Depression Rating Scale (HDRS). Classifier accuracy was defined as the percentage of classifier outputs during regulation blocks that matched the instructed mental state. Overall, session-specific, early- and late-training accuracy were expressed as percentages; changes in accuracy were expressed as percentage points; and the accuracy slope across Sessions 1-5 was expressed as percentage points per session. HDRS change was expressed in scale points and calculated as CE0-CE1, CE0-CE2, and CE1-CE2; positive values indicate symptom reduction. Pearson's r was used when both variables met normality assumptions; otherwise, Spearman's rho was used. Correlation coefficients are unitless and range from -1 to +1.

Correlations were calculated separately by arm between classifier performance and self-reported depressive symptom change measured with the Beck Depression Inventory-II (BDI-II). Classifier accuracy was defined as the percentage of classifier outputs during regulation blocks that matched the instructed mental state. Overall, session-specific, early- and late-training accuracy were expressed as percentages; changes in accuracy were expressed as percentage points; and the accuracy slope across Sessions 1-5 was expressed as percentage points per session. BDI-II change was expressed in scale points and calculated as CE0-CE1, CE0-CE2, and CE1-CE2; positive values indicate symptom reduction. Pearson's r was used when both variables met normality assumptions; otherwise, Spearman's rho was used. Correlation coefficients are unitless and range from -1 to +1.

The common spatial pattern filter matrix was inverted to derive spatial patterns for the 28 EEG channels. Participant-level patterns were normalized by Euclidean norm and averaged separately by arm, mental task, and session. The maps were descriptive; no inferential comparisons were planned.

Participants were repeatedly asked about perceived stress and comfort during the protocol. Intervention-related discomfort and adverse events were summarized by arm. This was a procedural tolerability assessment rather than a formally adjudicated adverse-event endpoint.