Matched-Input Estimates Differ in Sign Across Architectures: Auditing EEG Foundation Models on Motor Imagery

Pretrained EEG foundation models are proposed for clinical decoding, but whether reported gains transfer across populations or survive negative controls is unclear. We benchmark LaBraM, EEGMamba, CBraMod, REVE, LEAD, BENDR, and BIOT on five clinical tasks across four datasets. Primary analyses use frozen linear probes with subject-disjoint LOSO or grouped five-fold validation. Because CAUEEG releases no patient identifiers, it is evaluated at recording level with a patient-disjoint sensitivity. We challenge apparent gains using stronger classical comparators, label permutation, scrambled-label fine-tuning, and random-initialisation controls. In a matched 19-channel CAUEEG evaluation (Normal/MCI/Dementia; N = 1,187 recordings), classical features achieve 0.734 macro-AUROC versus 0.699 for BIOT, 0.669 for CBraMod, and 0.568 for REVE. A patient-disjoint sensitivity retains the classical-over-REVE ordering (0.717 versus 0.565). Dataset identity is decoded from frozen REVE embeddings at or near ceiling across Western-Korean and Western-Western pairs, including after PCA-50 and removal of line-frequency and amplitude-scale information. This establishes dataset membership, not a causal site or population effect. A matched random-initialised encoder exceeds pretrained REVE on CAUEEG (0.659 versus 0.570). On CHB-MIT cross-subject ictal detection (n = 23), REVE reaches 0.793, versus 0.739 for the best enhanced nonlinear comparator, 0.701 for random initialisation, and 0.505 for raw-signal random features. Because preprocessing removes absolute amplitude, this does not establish superiority over every plausible handcrafted baseline. Conclusions change materially after montage matching, patient-overlap checks, stronger comparators, and representation controls. We distill these checks into a reporting protocol for clinical EEG foundation-model studies.

Scaling EEG foundation models requires pooling data across heterogeneous electrode montages, a prerequisite both for larger pretraining corpora and for downstream deployment. We present the first systematic comparison of four channel adaptation methods (Conv1d projection, spherical spline interpolation (SSI), source-space decomposition, and Riemannian re-centering) across five pretrained EEG foundation models (5M--157M parameters), five downstream tasks, and two training regimes with 10--15 random seeds each. We find that rigid-montage models (BENDR, Neuro-GPT) require external adaptation, while flexible models (EEGPT, CBraMod) match or exceed it natively when fine-tuned but benefit from external methods under frozen-encoder deployment. A probe-SFT asymmetry exists: external adaptation can cause severe negative transfer during fine-tuning of flexible models. The optimal method is architecture-dependent (Conv1d for BENDR, SSI/Riemannian for Neuro-GPT, source-space decomposition for depression detection), and 5M-parameter CBraMod outperforms models up to 31$\times$ larger on 4/5 datasets, consistent with independent findings that compact EEG-specific architectures can match larger models.

Electroencephalography (EEG) decoding for brain-computer interfaces (BCIs) faces a major challenge: substantial inter-subject variability limits effective cross-subject generalization. Consequently, practical systems still rely largely on subject-specific models trained from scratch and requiring individual recalibration. EEG foundation models have recently emerged as a promising alternative; however, even large pretrained models cannot simply be used as fixed feature extractors and still require additional adaptation before they can be reliably applied to downstream tasks. In this work, we address this challenge through targeted adaptation strategies. Building on recent EEG foundation models such as REVE, LaBraM, and LUNA, we examine the impact of different low-rank adaptation strategies on motor imagery classification. We propose a framework that structurally decouples subject-invariant knowledge from subject-specific neural signatures: the low-rank update at each adapted layer is split into a Global adapter, trained jointly across all subjects, and Subject-Specific adapters, each absorbing individual variability. To assess the contribution of each path, we compare three adaptation strategies: (i) subject-specific LoRA (ii) global LoRA and (iii) stacked LoRA, combining both Global and Subject Specific adapters. Experiments on BCI Competition IV-2a, PhysioNet Motor Imagery, and the clinical Zuo2025 benchmark show that Stacked LoRA effectively mitigates inter-subject variability, achieving the best accuracy in the large majority of backbone and dataset combinations. Our analysis further reveals that the optimal balance between the global and subject-specific paths depends on the target population: a shared adapter is sufficient for large, diverse cohorts, whereas subject-specific adaptation is decisive under the high inter-session variability of clinical recordings.