EEG Phase-Resetting Index Tracks Depression Severity in 276 Untreated Adults

A 14-channel consumer-grade scalp EEG system has yielded a cross-validated prediction of self-reported depressive symptom severity at r = 0.620 in 276 untreated adults, with roughly 80% accuracy in separating moderate-to-severe from low-symptom individuals. The study, published in Nature Mental Health by Masataka Morita and Hiroshi Konishi of the University of Tsukuba, appears in the journal's 2026 volume (DOI: 10.1038/s44220-026-00757-8).

The work rests on a new measure called phase-resetting rate modulation, or PRRM: a source-locked, time-lagged descriptor of how phase-resetting events at one electrode modulate the phase-resetting rate at another. The signal lives not in any single channel but in an integrated index aggregating multiple individually weak features across the upper-alpha (10.5–12.5 Hz) and low-gamma (33–44 Hz) bands.

What did the cohort look like?

All 276 participants — 139 males, 137 females, aged 18–59 (mean 43.25 ± 10.49 years) — were medication-free and had received no psychiatric treatment within the past year. Recruitment drew on roughly 4,200 screened applications; six participants were excluded for inconsistent questionnaire responses before recording.

Symptom severity was graded with the Zung Self-Rating Depression Scale (SDS) as the primary measure and the PHQ-9 as a complement. The two scales correlated strongly overall (r = 0.922), but concordance dropped to r = 0.663 in the low-severity range (SDS ≤ 39, n = 114), where the PHQ-9 showed a pronounced floor effect — 59 participants, 21.4% of the sample, scored zero. The final grouping: low (SDS ≤ 39, n = 114), border (40–47, n = 46), high (SDS ≥ 48, n = 116).

How was the EEG recorded?

Recording used an Emotiv EPOC X wireless headset: 14 electrodes positioned per the extended 10–20 system (AF3, F3, F7, FC5, T7, P7, O1 and right-hemisphere homologues), sampled at 128 Hz with a 0.2–45 Hz bandwidth, CMS reference at P3 and DRL at P4. Each participant completed four 2.5-minute eyes-closed resting sessions — 10 minutes total. Of 1,104 sessions, only 12 failed quality screening (contact quality ≤ 70%, sustained 0% contact, or duration under 2 min 25 s).

Notably, the team applied no ICA or artifact-subspace reconstruction. Artifact control relied on acquisition-stage monitoring, session repetition and conservative session-level screening, a deliberate choice given the sparse 14-channel montage and unvalidated effects of decomposition on phase-reset detection.

What did the model achieve?

Standard features failed first. Relative alpha power and frontal alpha asymmetry showed no meaningful cross-validated prediction of SDS scores in this cohort; raw phase-resetting rate showed only weak association with substantial inter-individual variability.

PRRM changed that picture. After converting to relative PRRM (subtracting each target's mean across sources), permutation tests across 182 directed electrode pairs identified 13 features at uncorrected P < 0.01 — versus 3.64 expected by chance — with low inter-feature redundancy (mean absolute correlation 0.064). A descriptive full-sample integrated index correlated with SDS at r = 0.726 (95% CI 0.66–0.78), robust to age and sex adjustment.

Under stratified k-fold cross-validation, with feature selection, standardization, clipping and regression fitted strictly within training folds:

• At k = 46: r = 0.620 (95% CI 0.542–0.688), mean absolute error 8.64 SDS points.
• Binary discrimination (high vs. low, border excluded, threshold 43.5): accuracy 0.791 (182/230), sensitivity 0.767, specificity 0.816.
• At leave-one-out (k = 276): r = 0.679, MAE 8.02, accuracy 0.817.

The low-gamma range contributed more strongly than upper alpha; lower-alpha and beta ranges showed little predictive utility. Only one electrode pair (P7 → O1) survived Bonferroni or FDR correction, and its standalone correlation was modest (r = 0.301) — underscoring why the aggregation strategy matters.

Why it matters for the reading room

This is proof of principle, not a clinically ready tool. Symptoms were self-reported, not confirmed by structured clinical interview; the design was cross-sectional and internally validated only. The sparse 14-channel coverage, 128 Hz sampling and 45 Hz bandwidth limit spatial localization and may under-capture gamma-band dynamics — PRRM patterns cannot be assigned to specific anatomical generators, and the authors explicitly frame PRRM as an event-locked descriptor, not a causal connectivity measure.

Practitioners should also note the disclosure: the University of Tsukuba has filed patent applications covering the PRR and PRRM methods, with Morita and Konishi as named inventors. The study was funded by JSPS KAKENHI grant 23K17457; de-identified EEG data and MATLAB preprocessing code are available via Figshare (DOI: 10.6084/m9.figshare.31867723).

What comes next?

The authors call for external validation in clinically characterized cohorts, longitudinal testing of whether the index tracks symptom trajectories and treatment change, and denser, higher-bandwidth systems that could shorten recording duration or improve stability. If the framework generalizes beyond depressive severity, PRRM may offer a physiologically interpretable route to characterizing internal brain states across a wider range of mental-health conditions.