Does keeping template evidence improve EEG microstate sensitivity to Alzheimer's?

The hard label discards evidence strength

Conventional EEG microstate analysis reduces each scalp voltage pattern to a winner-take-all label: the topography is assigned to whichever template it most resembles, and the resulting label sequence is summarised by duration, occurrence, coverage, transitions and symbolic complexity. The paper's starting observation is that this readout cannot distinguish a time point with one clearly dominant template from a time point where two templates are nearly tied. Once the evidence profile is collapsed to its argmax, the winner's strength, its margin over competitors, and whether the evidence was mixed or weak are no longer recoverable.

The authors treat the hard label sequence as a compressed readout of a richer object they call a template evidence trajectory. At each sampled Global Field Power (GFP) peak — the local maxima of the across-channel standard deviation of the alpha-band signal, used as the standard high signal-to-noise sampling point in microstate work — the trajectory keeps the similarity to every template, not just the best one. The question the study asks is whether the discarded evidence can support matched, evidence-aware versions of the classical duration and complexity descriptors.

Two descriptors built from the same evidence

The framework derives two extensions. High-evidence episode duration and episode rate threshold the trajectory so that only samples with strong evidence for a state count, then measure whether that strong evidence clusters into sustained episodes or fragments into short bursts. Because the threshold is a within-state quantile, the marginal amount of high-evidence samples is fixed by construction; the descriptor therefore does not estimate absolute occupancy, but the temporal organisation of confidence.

Null-state Lempel-Ziv complexity (LZC) targets the symbolic sequence instead. Ordinary LZC treats every hard label as a meaningful symbol even when no template explains the topography well. Null-state LZC assigns low-evidence samples to a dedicated null symbol, so periods of insufficient evidence enter the complexity computation explicitly rather than being counted as ordinary state visits.

Both extensions are computed from the same template-similarity matrix as the conventional descriptors, which is what makes the comparison matched: for each state model, the authors derive a hard label sequence and a trajectory from identical inputs, then compare the statistical effects of the paired descriptors.

Four datasets and eight state models

The evaluation spans four eyes-closed resting-state EEG datasets covering two validation settings. For Alzheimer's disease versus healthy control contrasts, the study used OpenNeuro ds004504 (36 AD and 29 healthy control participants, with the frontotemporal dementia group excluded) and the CAUEEG cohort (228 AD and 272 healthy control participants after restricting labels to early- or late-onset AD and normal controls). For age-related variation, it used OpenNeuro ds005385 from the Dortmund Vital Study (608 healthy adults aged 20-70) and TD-BRAIN (179 healthy adults aged 18 and over).

Each dataset was analysed under eight state model constructions: K-means, AAHC and HMM at K = 4 and K = 7, plus subject-specific adaptive Leiden and Infomap communities. That yields 32 dataset-model cases. The primary comparisons use global descriptors — global mean duration and LZC — rather than state-specific measures, because global descriptors are not tied to a particular template identity and therefore remain comparable across fixed-K settings and adaptive models with subject-specific state identities.

As a sanity check on the conventional branch, group-level K-means templates at K = 4 corresponded to the canonical A-D microstate patterns, with global explained variance between 0.66 and 0.73 across datasets; at K = 7 the repertoire was finer-grained and explained variance rose to 0.72-0.78.

Duration effects grew in every case

The headline result concerns effect-size ratios, defined as the absolute trajectory effect divided by the absolute matched hard-label effect. For the duration family, the mean ratio ranged from 1.34 to 1.44 across tested threshold and sharpening combinations. The largest mean ratio, 1.44, occurred at the representative setting, where the trajectory descriptor exceeded the hard-label descriptor in all 32 cases. Neighbouring settings produced ratios of similar size, so the advantage was not an artefact of one threshold choice.

Individual cases show what the ratio means in practice. In ds004504 with K-means at K = 4, the AD-versus-control duration effect moved from d = -0.871 for the hard label to d = -1.778 for the trajectory descriptor. In CAUEEG the same comparison moved from d = -0.486 to d = -0.655. In the age datasets, correlations strengthened from r = -0.172 to r = -0.252 in ds005385 and from r = -0.228 to r = -0.345 in TD-BRAIN, again for K-means at K = 4.

The authors are explicit that the direction of these effects is not the contribution. Shorter global duration and higher symbolic complexity appeared in AD and with increasing age in both AD datasets and both ageing datasets, and the evidence-aware descriptors amplified differences already present in the matched hard-label readout rather than imposing a new physiological direction.

Complexity gains are smaller and model-dependent

Null-state LZC also favoured the evidence-aware readout, but less uniformly. Across tested parameters the mean effect-size ratio ranged from 1.13 to 1.23. At the highlighted setting, null-state LZC exceeded hard-label LZC in 29 of 32 cases, and the maximum mean ratio of 1.23 was reached at several parameter settings rather than one.

Case-level numbers follow the same pattern of smaller but real gains. In ds004504 with AAHC at K = 4, the LZC effect rose from d = 1.147 to d = 1.824; in CAUEEG with the same model it rose from d = 0.748 to d = 0.919. The benefit was most consistent for constrained fixed-state models, where every GFP peak must be assigned to a limited template repertoire. For adaptive Leiden and Infomap models the incremental gain was attenuated, because ordinary hard-label LZC was already strong in some datasets.

This asymmetry matters for anyone deciding whether to adopt the framework. The duration extension is the robust part of the result; the complexity extension is a secondary gain whose size depends on how constrained the state model is.

Classification gains are modest but consistent

The authors then asked whether descriptor differences translate into subject-level prediction. Hard-label and trajectory feature sets each contained 16 descriptors, and a combined model concatenated all 32. Evaluation used repeated 5-fold outer cross-validation, with AD-versus-control tasks for the two AD datasets and a young-versus-old contrast (35 years or younger versus 60 years or older) for the two ageing datasets.

In ds004504, balanced accuracy rose from 0.806 for hard-label features to 0.828 for trajectory features and 0.839 for the combined model, while ROC-AUC was highest for the combined model at 0.888 versus 0.862 for hard labels. In CAUEEG, balanced accuracy rose from 0.704 to 0.729 and 0.739, and ROC-AUC from 0.778 to 0.790 and 0.814. In ds005385, balanced accuracy rose from 0.689 to 0.696 and 0.703, with the best ROC-AUC from trajectory features at 0.777. In TD-BRAIN, trajectory features produced the largest improvement, with balanced accuracy rising from 0.701 to 0.752 and ROC-AUC from 0.800 to 0.831.

Across datasets, the best balanced accuracy came from either the trajectory or the combined feature set. The gains are consistent in direction but small relative to the fold-to-fold standard deviations, so they support the descriptors as a useful addition rather than a replacement for the conventional feature set.

What the evidence-aware readout does not replace

The paper is careful about the scope of its claim. High-evidence episode metrics use within-state quantile thresholds, so they do not estimate absolute state occupancy; the authors state that conventional coverage remains a direct and interpretable measure of winner-state prevalence and should be retained. The evidence-aware descriptors are positioned as confidence-conditioned extensions of duration- and complexity-based readouts, not substitutes for the classical parameter set.

The signed effects are also described as pipeline-specific and hypothesis-generating. The authors note that replication would require varying preprocessing, template-fitting strategies, smoothing settings and state-number choices on the same datasets while reporting both signed and absolute effects, to determine whether shorter duration and higher LZC reflect a robust feature of particular AD or ageing cohorts, a consequence of alpha-band GFP peak dynamics, or an interaction between disease effects and state model construction. Analyses were restricted to eyes-closed resting-state alpha-band activity, with no cross-dataset channel interpolation or spatial harmonisation, and all topographic analyses were performed in each dataset's native montage.

Where the framework earns its place

The practical priority that emerges from the 32-case comparison is a clear ordering. If the goal is sensitivity to AD- or age-related variation in global topographic state dynamics, the high-evidence episode duration descriptor is the component to adopt first: it produced a larger effect than its matched hard label in every case at the representative setting and in at least 30 of 32 cases across the broader grid, with mean ratios of 1.34 to 1.44. Null-state LZC is worth adding where the state model is a constrained fixed-K construction, since its advantage shrank for adaptive Leiden and Infomap communities.

The classification results set the expectation for downstream use: trajectory or combined features were best or near-best in balanced accuracy on all four datasets, but the increments over hard labels were small. The framework's value therefore lies in sharpening the statistical contrast that microstate descriptors already detect, while remaining compatible with conventional analysis — not in replacing coverage, state-specific measures, or the hard-label parameter set.

Sources used in this answer