EEG Measures Hold Steady Across 43 Sites in AMP SCZ Reliability Test

Scalp EEG measures collected from 654 clinical high risk (CHR) participants across 43 sites on five continents held good-to-excellent test-retest reliability over roughly two months, according to interim results from the Accelerating Medicines Partnership® Schizophrenia (AMP SCZ) program published June 6 in Schizophrenia.

The interim analysis, led by Daniel H. Mathalon of the University of California San Francisco, covers 654 CHR individuals and 87 community controls who completed baseline and 2-month EEG sessions. Every site used an identically configured BrainProducts actiCHamp 64-channel active electrode system, referenced to FCz, with impedances kept below 25 kOhms where possible. The full recording session lasted 57 minutes and 23 seconds.

The stakes are straightforward. Conversion to psychosis in CHR cohorts has fallen from the 40–50% rates of early studies to roughly 19–25% over 2–3 years in recent meta-analyses. Without measures that predict who will convert, who will remit, and who will persist in the risk state, trials of preventive treatments remain inefficient.

What did the protocol include?

The EEG team selected four paradigms:

• A double-deviant mismatch negativity (MMN) task: 90% standard tones at 633 Hz, 10% pitch-plus-duration deviants at 1000 Hz, 3,200 tones over five runs, scored at a six-electrode fronto-central cluster peaking at FCz at 126 ms.
• Auditory and visual three-stimulus oddball tasks yielding target P3b and novel P3a, scored over centro-parietal and fronto-central clusters respectively.
• A 40-Hz auditory steady state response paradigm: 150 click trains of 500 ms, yielding inter-trial coherence, evoked power and total power in the 38–42 Hz band.
• Resting eyes-open and eyes-closed EEG, 185 seconds each, scored as power spectral density in delta, theta, alpha, beta and gamma bands.

How reliable were the measures?

Generalizability (G) coefficients, a form of intraclass correlation, were mostly good (0.6–0.75) to excellent (0.75–1.0) for MMN, all four P300 subcomponents, the three 40-Hz ASSR measures and most resting bands. Only resting gamma power fell into the fair range, a finding the authors attribute to the lower signal-to-noise ratio of gamma oscillations.

Because the median retest interval ran 10–11 weeks and true clinical change can inflate apparent error, the authors call these conservative estimates. Of 2,347 sessions rated for quality, 93.8% scored good or excellent; only 0.6% failed outright.

Amplitude stability told a more textured story. Auditory and visual target P3b, visual novel P3a and MMN all declined significantly from baseline to follow-up in both groups, consistent with habituation. Auditory novel P3a, curiously, did not budge. The P3b declines reflected steeper within-session habituation at the second visit: run-1 amplitudes for auditory P3b, visual P3a and MMN had not fully recovered their baseline values two months later.

What does this change for the reading room and the lab?

The practical question for anyone running repeated ERP sessions is whether amplitude drops signal pathology or practice. Here, task performance stayed high at both sessions — median target miss rates below 4% — so the declines reflect habituation, not deteriorating attention. Trial designs using these measures as target-engagement endpoints should expect and model that drift. Researchers planning longitudinal oddball work should also note the habituation pattern: first-run amplitudes are the least contaminated.

The harmonization effort deserves its own mention. Identical leased hardware, air-gapped recording laptops, translated technician scripts, remote certification of technicians, weekly two-hour video review calls, and a web-based quality control dashboard with automated PREP, FASTER and ICLabel processing produced data comparable across continents and languages. The SOP manual is publicly downloadable at ampscz.org.

These remain investigational research measures, not FDA-cleared diagnostics, and group differences between CHR and controls were deliberately not tested in this interim sample. All data are being released through the NIMH Data Archive at nda.nih.gov/ampscz, and the complete cohort — with group contrasts, clinical outcome predictions and age-normative modeling of the developmental span from age 12 upward — is the promised next installment.