When the same EEG biomarker replicates in Ireland but not in the Netherlands

What the work claims

The authors set out to validate a previously reported EEG signature of ALS: reduced beta-band event-related desynchronization (ERD) and attenuated post-movement beta event-related synchronization (ERS) during a sustained-attention-to-response task (SART). They recorded 128-channel EEG in a Dutch cohort (63 people with ALS, 64 controls) and an Irish cohort (36 ALS, 36 controls), applied an automated preprocessing and time-frequency pipeline, and tested whether the group differences reproduced across sites.¹

The study reports three things. First, the SART-elicited theta, alpha, and beta ERSP patterns in controls were highly consistent across the two centres. Second, the Irish ALS cohort showed reduced beta-band ERD and ERS compared with Irish controls, broadly replicating the earlier single-site finding. Third, the Dutch ALS cohort showed no significant beta-band group differences at all. Behavioural differences between cohorts could not fully explain the divergence.

How it works

The task is a Go/NoGo digit-detection paradigm. Digits appear for 250 ms with an inter-stimulus interval of 1120 to 1220 ms; the digit "3" is the NoGo target. Participants completed three or four 252-trial blocks. EEG was acquired with a BioSemi Active Two 128-channel system sampled at 512 Hz, hardware low-pass filtered at 104 Hz to prevent aliasing.¹

Preprocessing was deliberately automated and identical at both sites: a 0.3 Hz dual-pass fourth-order Butterworth high-pass filter, resampling to 256 Hz, detection and spline interpolation of flat or noisy channels, common-average re-referencing, ICA-based removal of ocular, muscle, and cardiac components, and a final 60 Hz dual-pass fourth-order Butterworth low-pass filter. Time-frequency decomposition used complex Morlet wavelets from 1 to 50 Hz. The analysis focused on frontoparietal electrodes Fz, Cz, and Pz, with baseline correction from -200 ms to stimulus onset.

The key readouts were baseline-normalized post-stimulus power decreases (ERD) and increases (ERS) in predefined time-frequency windows. Group discrimination was quantified with area-under-the-ROC-curve (AUROC) values and empirical-Bayesian posterior probabilities with false-discovery-rate correction. Supplementary ROC analyses report Go beta-band ERS at Fz in the Irish cohort achieved an AUC of 0.780 (95% CI 0.672 to 0.889) with 81% sensitivity and 69% specificity at the optimal operating point; the equivalent Dutch measure reached only 0.620 with 51% sensitivity and 83% specificity.¹

Where a skeptic should push

The strongest caveat is sample heterogeneity. The Dutch and Irish ALS cohorts differed in King's clinical stage distribution (p = 0.03), riluzole use (86.1% Irish versus 69.8% Dutch, p = 0.02), and education level. The Irish cohort also had more trials available for time-frequency analysis (median 128 versus 118, p = 0.01). The authors ran linear models with King's stage, riluzole, country, and education as predictors and found that none of these factors significantly explained the beta ERSP differences between centres, but that is not the same as proving the acquisition chain was innocent.

A second concern is response-speed confounding. The Irish ALS group had slower Go-trial reaction times than the Dutch ALS group (all p = 0.03), and greater beta ERS was associated with faster responses across both centres (p < 0.011). Beta ERS is a post-movement rebound; if reaction time varies between cohorts, a stimulus-locked analysis will smear or shift the rebound window. The study checked response-locked epochs and reports similar topographies, but the reported group discrimination is still built on stimulus-locked windows.

Third, the effect sizes are modest. Even in the Irish cohort the best discriminative feature, Go beta-band ERS at Fz, yielded an AUC of 0.78. That is clinically interesting but far from a standalone diagnostic test. The Dutch cohort's AUCs for the same features hovered around 0.54 to 0.63. The paper therefore demonstrates replicable control dynamics and a promising Irish replication, not a validated cross-site biomarker.

What this means for MEA biomarker reproducibility

The most useful reading for microelectrode array hardware is not about ALS itself. It is that a rigorously standardized protocol, identical amplifiers, identical software, and identical preprocessing still produced a disease-related electrophysiological signature in one site and not in another. If that can happen with 128-channel scalp EEG, it can certainly happen with a 26,000-channel CMOS-MEA recording from a living organoid.

The mechanism that matters. Beta ERS in this task is a motor-related rebound. The authors show that its amplitude covaries with reaction speed and that slower responses in the Irish cohort plausibly shift the rebound. In MEA terms, the equivalent problem is trial locking. An organoid assay that stimulates at fixed intervals and averages responses to a stimulus clock assumes the biological response is time-invariant. If the culture's excitability drifts from plate to plate, or if the electrode-tissue impedance changes with temperature and medium composition, the rebound or burst timing shifts. Averaging across those shifts dilutes the effect exactly the way a fixed stimulus window would dilute a reaction-time-dependent beta rebound.

The opportunity. The study's automated pipeline is a model for what organoid MEA workflows should look like: every preprocessing step specified, version-controlled, and applied identically across sites. The validation against the older manual pipeline is also instructive: N2 and P3 peak amplitudes correlated with Spearman rho values of 0.93 to 0.94 between pipelines, but latencies correlated lower (0.79 to 0.83). Latency is where subtle filter, sampling, and referencing differences show up first. For MEA systems, that argues for publishing the full acquisition specification: amplifier input-referred noise, anti-aliasing corner, reference scheme, ADC resolution, and post-processing filters. If a biomarker is real, it should survive transparent reporting and independent replication; if it disappears under those conditions, it was never a biomarker.

The threat. The paper's partial replication is a reminder that site effects can be mistaken for treatment effects. In a drug-screening campaign, one might compare vehicle and compound on organoids recorded on different MEAs, in different incubators, or with different reference electrodes. A beta-ERS-like metric that is sensitive to response speed or trial count will vary with those hardware variables. Without a calibration phantom or a within-site control, a "rescue" of a disease phenotype could be an artifact of a quieter amplifier or a better ground. The honest conclusion here is that the burden of proof for MEA biomarkers must include cross-site replication on independently acquired data, not just larger n at a single site.

The instrumentation implication. BioSemi active electrodes reduce motion and impedance sensitivity, and the authors still saw residual site differences. High-density MEAs face an even harder version of the same problem: thousands of tiny electrodes, each with its own impedance and noise, bathed in an electrolyte that changes over days. The path forward is to treat the electrode-tissue interface as a measured variable, not a constant. Closed-loop impedance monitoring, per-electrode calibration, and event-driven rather than purely stimulus-locked analysis are the engineering counterparts to the statistical corrections used here.

The bottom line

This is a well-executed replication study with an important negative result. It confirms that control ERSP dynamics are robust across independent EEG labs, but it does not confirm that beta-band ERD/ERS is a cross-site ALS biomarker. The Irish replication and the Dutch non-replication together point to residual heterogeneity that automated preprocessing alone cannot remove.

For MEA hardware, the paper is a caution: a plausible neural signal can look solid in one acquisition setup and vanish in another. What would strengthen confidence is a follow-up that explicitly varies acquisition parameters (reference, filtering, sampling rate, electrode type) and shows that the biomarker survives. Until then, organoid MEA results should be reported with their full acquisition provenance, and cross-site replication should be treated as a required gate, not an optional extra.