Three EEG Measures Survive a Pre-registered ALS Validation Filter
A pre-registered three-criterion framework narrowed a long list of candidate EEG biomarkers in amyotrophic lateral sclerosis to three measures with the strongest translational profile: gamma power, theta power, and spectral entropy. The study was built around a problem that has slowed EEG biomarker development for years—signals may look persuasive in a single cohort, yet fail when moved across settings, devices, or patient groups.
Investigators did not treat any single feature as a biomarker simply because it separated ALS from healthy participants. Each candidate had to clear three hurdles: cross-sectional discrimination, within-subject reliability, and cross-cohort effect direction consistency. Only features that met all three pre-specified requirements were allowed through the final screen. Three EEG features did so. That is a far tighter result than the broad biomarker claims common in neural engineering, and it comes with a built-in measure of caution.
The analysis drew on 18 EEG features across three independent cohorts: EEGET-ALS, which included six ALS patients and 170 healthy controls with home recordings over 17 days; UCL/Penn State, which included eight ALS patients with clinic recordings over one to two months; and a healthy adult test-retest reference cohort of 60 participants recorded across three sessions over one month. The design choice mattered because reproducibility was tested not in a single sample, but across distinct recording contexts that differed in ways likely to challenge any unstable signal.
Reliability Separated the Durable Signals from the Fragile Ones
Cross-sectional signal alone was not enough. Each feature had to show a Cohen’s d greater than 0.5, then demonstrate within-subject reliability with ICC(3,1) greater than 0.6 in both ALS cohorts independently, and finally preserve effect direction across cohorts. This approach mirrors the logic of translational biomarker work in other fields: a useful marker must be detectable, repeatable, and directionally stable before it can be considered for prospective deployment.
Gamma power, theta power, and spectral entropy cleared the full filter. Their emergence is clinically interesting because each reflects a different aspect of cortical dynamics. Power measures may track changes in oscillatory activity, while spectral entropy captures signal complexity and organization. In ALS, where cortical and motor network dysfunction can alter electrophysiologic patterns, these features are biologically plausible candidates. Yet the study did not present them as finished biomarkers; it presented them as the strongest candidates after a deliberately strict screen.
One result stood out sharply. A principal methodological finding was the extent to which feature reliability differed across cohorts relative to the disease-associated signal itself. Relative beta power showed an ICC of 0.09 in a home-recorded cohort but 0.93 in a clinic-recorded cohort. The gap is hard to ignore. It points to substantial cohort-associated variability, with protocol, hardware, setting, population, and preprocessing all changing together across datasets.
Cross-cohort Stability Was Real, but Sample Composition Still Mattered
The strongest features did not behave as though they were immune to sampling effects. Cross-cohort permutation testing showed that none of 1,000 permutations produced any feature in the three-cohort reliability intersection, yielding a Monte Carlo p value of 0.001. In plain clinical terms, the intersection observed in the real data was unlikely to be a random artifact of cohort shuffling.
Even so, bootstrap resampling revealed fragility in the composition of the selected set. The median count was 1, with a 95% confidence interval from 0 to 4, indicating that the exact makeup of the final biomarker set changed with sample composition. At a relaxed reliability threshold of ICC greater than 0.5, gamma power had the highest inclusion probability at 67%, followed by spectral entropy at 43% and theta power at 32%. At the primary pre-registered threshold of ICC greater than 0.6, those probabilities fell to 57%, 27%, and 19%, respectively.
That drop does not invalidate the framework. It reveals where biomarker development actually lives: in the space between promising effect sizes and imperfect reproducibility. ALS EEG research has repeatedly struggled at that boundary, especially when findings from one lab or one recording protocol fail to travel. Here, the protocol was not a minor detail. It was part of the cohort-level signal structure, and the data showed that clearly.
Protocol Effects Obscured the Disease Signal Across Sites
The most consequential methodological message is not that ALS has no EEG signature. It is that cross-cohort variability can overwhelm the apparent disease-related pattern when studies differ in recording environment and analytic pipeline. Because recording protocol, hardware, setting, population, and preprocessing all varied jointly between cohorts, the study could describe cohort-associated variability but not isolate a single causal protocol effect. That limitation is built into the design, not an afterthought.
This matters for how future EEG biomarker studies are built. A feature that performs well in one site may owe part of its apparent strength to local acquisition conditions rather than ALS biology. When the setting changes, the signal may weaken, invert, or lose reliability. The beta power example makes that problem tangible. A marker that is nearly unstable in one cohort and highly reproducible in another cannot be treated as a fixed biological truth without careful harmonization.
The framework itself is the main contribution. By pre-registering the screen, applying explicit thresholds, and quantifying uncertainty with both permutation testing and bootstrap resampling, the investigators created a more disciplined path for EEG biomarker triage. Gamma power, spectral entropy, and theta power now stand as priority candidates for prospective validation, not as finished diagnostic tools. The study also shows that the barrier to translation may be less about finding a signal than about proving that the signal survives across cohorts shaped by different technical and clinical conditions.
For ALS biomarker development, that is the central constraint. EEG markers will not reach clinical deployment on biological plausibility alone; they will need to prove they can survive the noise introduced by real-world recording differences, site-to-site heterogeneity, and analytic variation. This study makes that challenge visible in a way that is methodologically explicit and clinically relevant.