Clinical EEG models are usually trained on one electrode layout, with a fixed set of scalp positions in a fixed order. Once such a model is validated as a medical device, it is legally locked. A new hospital, though, records with its own cap, its own electrode count, its own reference scheme, and its own slightly misplaced electrodes. This proposal builds a small, physically constrained "montage adapter" that sits in front of the locked model. The adapter is a single linear map, derived from an electromagnetic head model, that converts the new site's recordings into the channels the locked model expects. Its few free parameters are electrode position offsets, per-channel gains, and the reference scheme. They are fitted using only unlabeled target recordings, by asking the head model to predict each held-out electrode from the others. No labels, no source data, and no weight change in the certified model are needed. Because the fit never looks at the model's predictions, it cannot be biased by the target site's class balance.
The opportunity comes from combining two results in the file. Adapt the Data, Not the Model showed that a trainable adapter in front of a frozen clinical predictor can beat end-to-end fine-tuning, for example by +14.1 AUCPR points on acute kidney injury prediction from MIMIC-IV to eICU. It also found that the adapter worked by correcting site-specific recording artifacts rather than by aligning distributions. That adapter, however, needs labeled target data and assumes identical channels at every site. Physics-informed and Unsupervised Riemannian Domain Adaptation showed that a static, Maxwell-equation-based "field interpolation" beats spline interpolation and common-channel selection when two EEG datasets share few electrodes. Its interpolation, though, is fixed in advance from nominal electrode positions. Our claim is that the residual cross-site gap after static field interpolation is largely a sensor-level gap: misplacement, gain, and reference. That gap can be closed label-free by fitting a handful of physical parameters, and doing so beats both static physics and the stronger statistical adaptation methods under compound montage-plus-label shift.
Core Claim
EEG hardware plays the role that scanner effects play in medical imaging. Each site imprints its own acquisition signature. For EEG, much of that signature has a known physical form. Scalp potentials are, to a good approximation, a linear mixture of cortical sources through a "lead field," the matrix that says how much each brain source contributes to each electrode. We model a target site's recording $x_t \in \mathbb{R}^{C}$ at time $t$ (with $C$ site-specific channels) as
$$x_t = D_g \, R_r \, G(p + \delta)\, s_t + n_t ,$$
where $G(p+\delta)$ is the lead field evaluated at nominal electrode positions $p$ plus unknown offsets $\delta$, $D_g$ is a diagonal matrix of per-channel gains $g$, $R_r$ applies the site's referencing (common average, linked mastoid, or bipolar, indexed by $r$), $s_t$ are the unknown source currents, and $n_t$ is sensor noise. Write $\theta = (\delta, g, r)$ for the site parameters. For any $\theta$, a regularized minimum-norm inverse followed by the template lead field gives a linear map $W(\theta)$ from the site's channels to the locked model's input channels. Static field interpolation is the special case $\delta = 0$, $g = 1$, with the reference assumed known.
We fit $\theta$ by leave-channel-out prediction. Each electrode's signal is predicted from all the others through the same inverse-forward model:
$$\hat\theta = \arg\min_\theta \sum_t \sum_{i=1}^{C} \frac{\big( x_{t,i} - [M_{-i}(\theta)\, x_{t,-i}]_i \big)^2}{\operatorname{var}(x_{\cdot,i})} + \lambda \, \|\delta\|^2 ,$$
where $M_{-i}(\theta)$ reconstructs channel $i$ from the remaining channels $x_{t,-i}$. The per-channel variance normalization removes the trivial gain rescaling, and $\lambda$ keeps offsets near plausible placement error, which is a few millimeters. The objective contains no class labels, no source-site data, and no predictions from the locked model. Label shift at the target therefore cannot bend it. That is exactly the failure Rectifying Geometric Misalignment documents for Riemannian re-centering at an imbalance ratio of 0.1. The output is one matrix with roughly a hundred interpretable parameters. A regulator or clinician can inspect it, which answers the auditability concern that learned black-box input transforms raise.
Relation to Prior Work
Three lines of work motivate the proposal. Adapt the Data, Not the Model supplies the deployment stance of a frozen predictor with adaptation in input space. It also supplies the diagnosis that the gains come from undoing operational artifacts. We keep the stance but replace its learned axial-attention adapter, which is label-supervised and assumes a fixed channel set, with a physics-parameterized one that is label-free and layout-flexible. Physics-informed and Unsupervised Riemannian Domain Adaptation supplies the method core, the lead-field interpolation to a template. We differ by treating the sensor geometry as unknown and estimable, and by feeding a locked nonlinear predictor rather than a tangent-space logistic regression. Test-Time Adaptation for EEG Foundation Models motivates the choice of objective. Entropy-based test-time adaptation (Tent, SHOT) collapsed on EEG, losing over 6 points on scalp-to-ear transfer. The optimization-free prototype method T3A was stable, and that paper explicitly leaves adaptation of the spatial front end open. We move adaptation into the sensor model and optimize a physical reconstruction target rather than the classifier's confidence.
Two works constrain the design. Rectifying Geometric Misalignment shows that any adaptation using target class statistics inherits label-shift bias; our objective avoids this by construction. Spatial Imputation Drives Cross-Domain Alignment (IMAC) maps electrodes onto a flat 9×10 grid with radial-basis interpolation and trains a Transformer imputer. It shows that masked-channel reconstruction is a useful self-supervised signal. We use the same signal, but inside a 3D physical model with a few parameters instead of a trained network, and against a locked downstream model rather than a jointly trained one.
Topology-agnostic foundation models are the natural competitor, not a building block. REVE encodes 3D electrode coordinates with Fourier features. LUNA compresses any channel set through learned queries. LUNA's own authors report that static coordinate encodings degrade when the target montage departs strongly from pre-training, as on SEED-V. Both assume the coordinates they are given are correct. Our adapter corrects the coordinates and gains themselves, so it is complementary. We test it both in front of a locked fixed-montage model and as a coordinate corrector for REVE.
Research Program
The work has three parts. First, we build the adapter. We use a standard template head model, compute lead fields on a dense cortical source grid, and derive $M_{-i}(\theta)$ and $W(\theta)$ in closed form for any $\theta$ through regularized minimum-norm operators. We fit $\theta$ by gradient descent on the leave-channel-out objective, with the discrete reference choice selected by the lowest residual. We ask how many unlabeled minutes are needed; our expectation is a few minutes, since the parameters number in the tens to low hundreds.
Second, we validate identifiability in simulation. We project realistic source activity through the head model, apply known offsets, gains, and reference changes, and test whether $\hat\theta$ recovers them. We also test whether errors in the template head relative to the individual's true anatomy get wrongly absorbed into $\delta$.
Third, we deploy against locked predictors. We train and then freeze standard predictors on a source dataset: a Riemannian tangent-space pipeline and EEGNet on BNCI2014_001's 22 channels, and a clinical abnormality classifier on TUAB's 10–20 montage. We never update them again. At each target site, only the adapter is fitted, on unlabeled target recordings.
Baselines
The naive reference point has two parts. The first is common-channel selection: keep only electrodes shared with the source montage and zero-fill the rest. This is what a locked model receives without effort. The second is spherical-spline interpolation to the source montage. Physics-informed and Unsupervised Riemannian Domain Adaptation reports both as clearly inferior to field interpolation when overlap is small. Running them confirms that our pipeline reproduces that ordering.
The strongest competitor on the core question is static field interpolation, combined with unsupervised Riemannian re-centering where the predictor is Riemannian, as in Physics-informed and Unsupervised Riemannian Domain Adaptation. It uses the same head model and the same template, and it is also label-free, so it isolates exactly our added ingredient. That paper reports it "significantly higher" than alternatives on Shin2017A, but no reported numeric accuracy for it was found on Lacuna. We will reproduce it on the same six MOABB datasets under leave-one-dataset-out evaluation. Beating it means a paired improvement in balanced accuracy across subjects on the low-overlap targets that is statistically reliable. We define that as a Wilcoxon signed-rank test at p<0.05 over subjects, with the improvement concentrated where sensor-parameter errors should matter most.
For the frozen deep-model setting, the strongest stable test-time method in the file is T3A, which recomputes classifier-head prototypes from unlabeled target features. Test-Time Adaptation for EEG Foundation Models reports that T3A gives a +18.9 point balanced-accuracy gain for REVE-Base on CHB-MIT and +4.8 points for CBraMod on scalp-to-ear transfer. We run T3A on the same frozen models, on the same unlabeled target windows, and with a matched hyperparameter search. We also run the combination of adapter followed by T3A, since the two act on different ends of the model. For label-shift experiments we add OSPDIM (Rectifying Geometric Misalignment), which reports more than 15% balanced-accuracy improvement over online re-centering on BNCI2014001 at imbalance ratio 0.1. Because OSPDIM needs a fixed montage, it runs after static field interpolation. As a label-using upper reference, we train INPUTADAPTER-style supervised adapters on labeled target data, extending the method from Adapt the Data, Not the Model to operate after static field interpolation. This tells us how much of the supervised gain a label-free physical fit recovers.
The ablations isolate the key ingredient. We fit offsets only, gains only, reference only, and all three. We also replace the physics-based leave-channel-out objective with the same parameters fitted by the locked model's prediction entropy, which tests whether the physical target, rather than just the parameterization, is what matters.
All methods receive the same unlabeled target data, the same frozen predictor, and the same budget for hyperparameter search. Hyperparameters are tuned on held-out source subjects only, never on target labels.
Why not just…?
Why not just use static field interpolation, since it already handles non-overlapping montages? This is the real threat. If nominal electrode positions are accurate and gains are uniform, our fitted $\theta$ will sit near zero and add nothing. The decisive experiment is the head-to-head on Shin2017A, Zhou2016, and PhysionetMI, whose montages overlap least with the other sets. If fitted parameters do not beat static interpolation there, we will report that the cross-site gap is not sensor-level. That would also partly contradict the artifact-correction reading of Adapt the Data, Not the Model for EEG, and it would be a useful negative result. Simulation tells us beforehand what effect size is detectable.
Why not just use a topology-agnostic foundation model like REVE and skip locked fixed-montage predictors entirely? Deployed and certified clinical EEG software is overwhelmingly fixed-montage today, and replacing it means re-certification. More substantively, REVE and LUNA consume the electrode coordinates they are given. LUNA itself attributes its degradation on unusual montages to coordinate encodings that do not transfer. We test this directly by feeding REVE-Base corrected coordinates and gain-rescaled signals from our fit, compared with nominal ones, on CHB-MIT and SleepEDF-78. SleepEDF-78 is the target where Test-Time Adaptation for EEG Foundation Models found even T3A gave only marginal gains.
Why not just apply T3A or another prototype method after interpolation? T3A corrects where class centroids lie in the model's feature space. It cannot undo a reference mismatch or a mis-scaled channel before the features are computed. It also re-estimates class statistics from target data, so it should degrade under label shift in the way re-centering does. Our label-shift experiment tests this, and the adapter-then-T3A combination shows whether the two gains add up or overlap.
Why not just fit the adapter by minimizing the locked model's prediction entropy? That is our entropy-objective ablation. Prior evidence predicts it will collapse or chase the majority class: Test-Time Adaptation for EEG Foundation Models shows entropy methods losing over 6 points, and Rectifying Geometric Misalignment shows entropy-style centering failing under imbalance. The few-parameter physical constraint might tame entropy minimization, and if it does, we report that honestly.
Proposed Experiments
The simulation study comes first. We measure recovery error of offsets (in millimeters), gains, and reference choice as a function of recording length and channel count, down to the 3-channel BNCI2014_004 layout. That layout is likely under-determined, and we expect the fit to fail there.
The main experiment is leave-one-dataset-out motor imagery transfer across BNCI2014_001, BNCI2014_004, PhysionetMI, Shin2017A, Weibo2014, and Zhou2016. It uses frozen Riemannian and EEGNet predictors trained on the source, and reports balanced accuracy and AUC per target subject. This matches the protocol of Physics-informed and Unsupervised Riemannian Domain Adaptation, so the comparison with static interpolation is like-for-like. The MOABB benchmark (The largest EEG-based BCI reproducibility study) found within-dataset motor-imagery AUCs frequently above 0.80 for Riemannian pipelines. That figure serves as a rough ceiling for how much of the within-site performance cross-site adaptation recovers.
The compound-shift test repeats the main experiment with target classes subsampled to imbalance ratios of 1.0, 0.5, and 0.1, following Rectifying Geometric Misalignment. The prediction is specific: the adapter's accuracy stays flat as imbalance grows, while re-centering and T3A degrade. If the adapter degrades too, the label-shift-immunity claim is false.
The clinical test freezes a TUAB-trained abnormality or event model on the referential 10–20 montage and deploys it on CHB-MIT, a pediatric dataset recorded with bipolar references. Here the reference parameter $r$ and pediatric head size, absorbed partly into $\delta$, are the dominant effects. We also run the scalp-to-ear transfer from the NeuroAdapt-Bench setup. For the ear data we expect the template head model to be least accurate, so it serves as a stress test that marks the method's limits.
Across all settings we report the fitted parameters themselves: whether estimated offsets are consistent across subjects from the same site, as a shared cap would imply, and whether estimated reference choices match the documented recording protocols. This is the auditability check. It is the property a physically parameterized adapter offers and a learned one does not.
References
- Adapt the Data, Not the Model: Input-Space Adaptation for Frozen Time-Series Predictors
- Physics-Informed and Unsupervised Riemannian Domain Adaptation for Machine Learning on Heterogeneous EEG Datasets
- Rectifying Geometric Misalignment: Online Source-Free Adaptation for Class-Imbalanced EEG
- Test-Time Adaptation for EEG Foundation Models: A Systematic Study under Real-World Distribution Shifts
- Spatial Imputation Drives Cross-Domain Alignment for EEG Classification
- REVE: A Foundation Model for EEG - Adapting to Any Setup with Large-Scale Pretraining on 25,000 Subjects
- LUNA: Efficient and Topology-Agnostic Foundation Model for EEG Signal Analysis
- The largest EEG-based BCI reproducibility study for open science: the MOABB benchmark