Brain signals recorded at different sites look different. Skull shape, electrode placement, amplifiers and task protocols all shift the statistics of electroencephalography (EEG, scalp recordings of the brain's electrical activity). Classifiers trained on one group of people therefore degrade on the next. Federated learning trains one model across sites without pooling their recordings, and it inherits this problem. Our proposal is simple. Before any model is trained, each site computes a large set of classical hand-crafted EEG features, such as band power, Hjorth parameters, wavelet subband statistics, entropies and covariance-based spatial features. The sites then jointly keep only the features whose class-separating effect is both large and consistent across sites. Features that separate the classes strongly at some sites and not at others are dropped, even when their pooled effect looks impressive.
The selection step borrows the logic of random-effects meta-analysis, the statistical method for combining effects across studies that disagree. Each site reports a per-feature effect size. The federation estimates both the average effect and the between-site spread, and ranks features by how reliably their effect would carry over to a new, unseen site. Everything the server needs can be written as sums over sites. It can therefore be computed in one communication round under secure aggregation, which is a protocol that reveals only the sum of site contributions and never any individual site's numbers. The claim we test is that this "consistency-aware" selection, and not multi-domain features as such, is what makes hand-crafted features beat end-to-end deep networks for unseen subjects and sites in federated EEG decoding.
Core Claim
The closest existing work falls into two groups. The first is federated EEG models that learn their own representation. Federated Transfer Learning for EEG Signal Classification uses Riemannian covariance layers with a Maximum Mean Discrepancy (MMD) alignment loss, a penalty that pulls the feature distributions of different subjects toward each other, and it tests only binary motor imagery. Federated deep transfer learning for EEG decoding using multiple BCI tasks and the 'Sandwich' meta-framework use convolutional front-ends at each site with shared middle layers. mixEEG applies mixup-style augmentation to raw inputs. None of these asks which interpretable features are stable across sites.
The second group is feature selection itself. SeiFS performs cryptographically secure federated feature selection on non-EEG tabular data, and it ranks features by pooled discriminability. The EEG selectors of Li et al. (L1-penalized), FBCSP (mutual-information-based) and Wang et al. (random-forest-based) are all centralized and also rank by pooled discriminability.
Our method differs from all of these in its selection criterion. It penalizes between-site heterogeneity of the effect. It does not reward pooled separation alone. This criterion also offers a principled resolution of an open contradiction in the file. Li et al. found that simple linear features, especially beta-band Hjorth mobility, generalize best across subjects on SEED. Wang et al. found that non-linear entropies of wavelet subbands give near-perfect epilepsy detection, but they did not test on unseen subjects. Our hypothesis is that neither feature family is intrinsically "more robust." The two papers measured different things: Li et al. measured cross-subject consistency, while Wang et al. measured within-cohort separability. A heterogeneity-aware selector should recover Li et al.'s preference on SEED without our imposing it.
Research Program
Each site extracts the same fixed feature bank from every trial. Where site montages differ, the bank is computed on the common channel set, following the 17-channel convention of MF-SCSN, plus scalp-region averages. The bank has six feature families: time-domain statistics and Hjorth mobility and complexity; band power in delta through gamma; db4 wavelet subband statistics; approximate, sample and fuzzy entropy of the subbands; filter-bank log-variance features; and tangent-space coordinates of spatial covariance matrices. Each site z-scores its own features using its own data, which is a cheap correction for per-site baseline power shifts.
For feature $j$ and each one-versus-rest class contrast, site $k$ computes a standardized mean difference $d_{jk}$ and its sampling variance $v_{jk}$. Standard DerSimonian–Laird random-effects estimation then gives a between-site variance $\tau_j^2$ and a pooled effect $\hat\mu_j$ with standard error $\mathrm{SE}_j$. We score each feature by how far its effect is expected to stay from zero at a new site:
$$ s_j = \frac{|\hat\mu_j|}{\sqrt{\tau_j^2 + \mathrm{SE}_j^2}}. $$
The denominator is the width of the random-effects prediction interval. Its form comes from treating a new site's true effect as a draw from $\mathcal{N}(\hat\mu_j, \tau_j^2)$. A pooled Fisher-style score keeps only the $\mathrm{SE}_j$ term, and that single term is the difference between our method and the obvious baseline. For multi-class tasks we average $s_j$ over contrasts.
The DerSimonian–Laird quantities are built from site-level sums: $\sum_k w_{jk}$, $\sum_k w_{jk} d_{jk}$ and $\sum_k w_{jk} d_{jk}^2$. The server therefore receives only aggregates under secure aggregation. The cost is one round of a few numbers per feature, which is negligible next to the repeated weight exchanges of the deep baselines.
A second pass adds redundancy pruning, in the spirit of the multicollinearity concern in Wang et al.. Pooled feature correlations are also sums of per-site cross-products, so they can be aggregated securely. For each correlated group, the pass keeps the member with the highest $s_j$. A small classifier is then trained by federated averaging (FedAvg) on the selected features: we use multinomial logistic regression and a two-layer network. We also report which feature families survive at each benchmark as a scientific output in its own right.
Baselines
The naive reference points come first. One is isolated local training: each site fits its own classifier on the full feature bank. Federated GNNs for EEG-Based Stroke Assessment shows the collaboration gap this leaves, with local MAE of 3.68–3.92 against 3.22–3.23 federated. The other is FedAvg on the full, unselected feature bank. Both use the same classifier, rounds and tuning budget as our method.
The strongest end-to-end competitors differ by benchmark. On SEED it is mixEEG, which reports domain-generalization federated accuracy rising from 77.0% under FedAvg to 86.3%. On BEETL it is MF-SCSN, which reports 51.2% mean accuracy against 48.2% for ShallowConvNet, together with the Sandwich Deep Set variants, which report up to 9% improvement over EEGInception. We did not find absolute numbers for the Sandwich variants in the file.
We re-run these methods ourselves where code allows, on identical leave-one-subject-out or held-out-site splits and with matched hyperparameter search. We also quote the published figures for context. Beating mixEEG means exceeding 86.3% on SEED domain generalization. Beating MF-SCSN means exceeding 51.2% on the BEETL test subjects, in both cases with a single communication round for selection.
The geometric competitor is FTL-style federated tangent-space classification, which we extend to four classes with a one-versus-rest head. On BCICIV-2a we quote the centralized ceilings of FBCSP, with kappa 0.569, and EEG-DG, with 81.79% accuracy, as upper references rather than federated targets. We found no reported federated result on Lacuna for BCICIV-2a as a standalone benchmark.
The ablation that isolates our key ingredient is federated selection by pooled effect alone, dropping $\tau_j^2$. This is essentially what a SeiFS-style or federated Fisher/mutual-information selector computes. It uses the same number of retained features, the same redundancy pruning and the same classifier. Two further ablations check that the gains are not about something else. One is a family-level oracle that keeps only Hjorth and band power, as Li et al. would suggest. The other is our full method without per-site z-scoring.
Why not just…?
Why not just select by pooled discriminability, as centralized EEG work and SeiFS do? This is the real threat. If between-site variance $\tau_j^2$ is small relative to the pooled effect for most features, the two rankings coincide and our method adds nothing. The proposal settles this directly through the pooled-effect ablation. It also makes a falsifiable prediction: the gap should grow with measured heterogeneity. It should be small on SEED, where all subjects are in one lab, and largest on BEETL, where sites are whole datasets with different hardware and paradigms. If the gap is flat across that range, we will report that pooled selection suffices.
Why not just train end-to-end on raw EEG and let the network find stable features? The file suggests raw-input FedAvg generalizes poorly to unseen subjects. mixEEG's FedAvg baseline reaches 43.6% on CHB-MIT and 77.0% on SEED, and FTL reports about 6% over EEGNet and DeepConvNet. However, mixEEG's augmented models are strong, so we treat them as the target to beat, not a straw man. A tie at far lower communication cost, with an interpretable feature list, would still be a useful outcome.
Why not just per-site normalization plus L1-regularized federated training on all features? L1 penalizes weak features, but it cannot distinguish a feature that is consistently moderate from one that is strong at three sites and absent at a fourth. The "full bank plus z-scoring" baseline trained with an L1 penalty measures exactly how much normalization and sparsity alone recover.
Proposed Experiments
On SEED, each subject is a client, and we evaluate by leave-one-subject-out. On DEAP we do the same, with Li et al.'s 59.06% centralized result as the reference. On BEETL, each source dataset is a site, and we test on the held-out CybathlonIC and Weibo2014 subjects. On BCICIV-2a, four-class motor imagery, subjects are clients. On CHB-MIT, patients are clients, which lets us test whether entropies survive selection on the task where Wang et al. found them decisive. Metrics are accuracy and Cohen's kappa, with the variance across held-out clients, plus communication in bytes.
The decisive checks are three. First, consistency-aware selection should beat pooled selection on unseen-site accuracy, with the margin increasing with the estimated average $\tau^2$ across benchmarks. Second, the surviving feature families should reproduce Li et al.'s Hjorth and beta-band preference on SEED while retaining subband entropies on CHB-MIT. That pattern would show the linear-versus-non-linear contradiction to be about cross-site consistency rather than feature type. Third, the selected low-dimensional model should match or beat mixEEG and MF-SCSN on their own benchmarks. A failure of the first check would retire the idea. A failure of only the third would leave a cheaper, interpretable method that trails deep models.
References
- Federated Transfer Learning for EEG Signal Classification
- Federated deep transfer learning for EEG decoding using multiple BCI tasks
- The ‘Sandwich’ meta-framework for architecture agnostic deep privacy-preserving transfer learning for non-invasive brainwave decoding
- mixEEG: Enhancing EEG Federated Learning for Cross-subject EEG Classification with Tailored mixup
- Secure and Lightweight Feature Selection for Horizontal Federated Learning
- Exploring EEG Features in Cross-Subject Emotion Recognition
- Filter Bank Common Spatial Pattern Algorithm on BCI Competition IV Datasets 2a and 2b
- An automated detection of epileptic seizures EEG using CNN classifier based on feature fusion with high accuracy
- Federated GNNs for EEG-Based Stroke Assessment
- EEG-DG: A Multi-Source Domain Generalization Framework for Motor Imagery EEG Classification