Paper deep dive

PRISM: Exploring Heterogeneous Pretrained EEG Foundation Model Transfer to Clinical Differential Diagnosis

Jeet Bandhu Lahiri, Parshva Runwal, Arvasu Kulkarni, Mahir Jain, Aditya Ray Mishra, Siddharth Panwar, Sandeep Singh

Year: 2026Venue: arXiv preprintArea: cs.LGType: PreprintEmbeddings: 41

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%

Last extracted: 7/20/2026, 4:22:11 AM

Summary

The paper introduces PRISM, a masked autoencoder-based EEG foundation model, to investigate how pretraining population diversity affects downstream clinical performance. Comparing a narrow-source corpus (TUH + PhysioNet) against a multi-source corpus (including South Asian data), the study finds that diverse pretraining yields more adaptable representations under fine-tuning, significantly outperforming narrow-source models on the novel task of distinguishing epilepsy from diagnostic mimickers (PNES, syncope) by 12.3 percentage points. The work also highlights that dataset count is a confounding variable and identifies systematic inconsistencies between EEG-Bench and EEG-FM-Bench that reverse model rankings.

Entities (11)

PRISM · model · 98%Epilepsy vs. Mimickers · task · 96%REVE · model · 95%EEG-Bench · benchmark · 94%EEG-FM-Bench · benchmark · 94%Psychogenic Non-Epileptic Seizures · condition · 93%PhysioNet · dataset · 92%TUH Corpus · dataset · 92%South Asian Clinical Recordings · dataset · 90%CBraMod · model · 88%LaBraM · model · 88%

Relation Signals (8)

PRISM → performstask → Epilepsy vs. Mimickers

confidence 96% · we introduce the first evaluation of an EEG foundation model on distinguishing epilepsy from its diagnostic mimickers

PRISM → usespretrainingdata → TUH Corpus

confidence 95% · D1 (Narrow-source): TUH Corpus [18] + PhysioNet Motor Imagery

PRISM → usespretrainingdata → South Asian Clinical Recordings

confidence 95% · D2 (Multi-source): D1 augmented with multi-center South Asian clinical recordings

EEG-Bench → contrastswith → EEG-FM-Bench

confidence 94% · systematic inconsistencies between EEG-Bench and EEG-FM-Bench reverse model rankings on identical datasets by up to 24 pp

PRISM → evaluatedon → EEG-Bench

confidence 92% · EEG-Bench and EEG-FM-Bench differ in six concrete methodological dimensions

PRISM → evaluatedon → EEG-FM-Bench

confidence 92% · EEG-Bench and EEG-FM-Bench differ in six concrete methodological dimensions

South Asian Clinical Recordings → augments → TUH Corpus

confidence 90% · D2 (Multi-source): D1 augmented with multi-center South Asian clinical recordings

PRISM → outperforms → REVE

confidence 90% · PRISM matches or outperforms REVE (92 datasets, 60,000+ hours) on the majority of tasks

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:EEG foundation models are typically pretrained on narrow-source clinical archives and evaluated on benchmarks from the same ecosystem, leaving unclear whether representations encode neural physiology or recording-distribution artifacts. We introduce PRISM (Population Representative Invariant Signal Model), a masked autoencoder ablated along two axes -- pretraining population and downstream adaptation -- with architecture and preprocessing fixed. We compare a narrow-source EU/US corpus (TUH + PhysioNet) against a geographically diverse pool augmented with multi-center South Asian clinical recordings across multiple EEG systems. Three findings emerge. First, narrow-source pretraining yields stronger linear probes on distribution-matched benchmarks, while diverse pretraining produces more adaptable representations under fine-tuning -- a trade-off invisible under single-protocol evaluation. Trained on three source corpora, PRISM matches or outperforms REVE (92 datasets, 60,000+ hours) on the majority of tasks, demonstrating that targeted diversity can substitute for indiscriminate scale and that dataset count is a confounding variable in model comparison. Second, on a clinically challenging and previously untested task -- distinguishing epilepsy from diagnostic mimickers via interictal EEG -- the diverse checkpoint outperforms the narrow-source checkpoint by +12.3 pp balanced accuracy, the largest gap across all evaluations. Third, systematic inconsistencies between EEG-Bench and EEG-FM-Bench reverse model rankings on identical datasets by up to 24 pp; we identify six concrete sources including split construction, checkpoint selection, segment length, and normalization, showing these factors compound non-additively.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

39,388 characters extracted from source content.

Expand or collapse full text

PRISM: Exploring Heterogeneous Pretrained EEG Foundation Model Transfer to Clinical Differential Diagnosis Jeet Bandhu Lahiri 1 , Parshva Runwal 2 , Arvasu Kulkarni 2 , Mahir Jain 2 , Aditya Ray Mishra 2 , Siddharth Panwar 1,2 , and Sandeep Singh 2 1 Indian Institute of Technology Mandi, India d23146@students.iitmandi.ac.in 2 NeuroDx, India parshva@neurodx.ai, arvasu@neurodx.ai, mahir@neurodx.ai, aditya.mishra@neurodx.ai, siddharth@neurodx.ai, sandeep@neurodx.ai Abstract. EEG foundation models are typically pretrained on narrow- source clinical archives and evaluated on benchmarks from the same ecosystem, leaving unclear whether representations encode neural physiol- ogy or recording-distribution artifacts. We introduce PRISM (Population- Representative Invariant Signal Model), a masked autoencoder ablated along two axes—pretraining population and downstream adaptation— with architecture and preprocessing fixed. We compare a narrow-source EU/US corpus (TUH + PhysioNet) against a geographically diverse pool augmented with multi-center South Asian clinical recordings across multi- ple EEG systems. Three findings emerge. First, narrow-source pretraining yields stronger linear probes on distribution-matched benchmarks, while diverse pretraining produces more adaptable representations under fine- tuning—a trade-off invisible under single-protocol evaluation. Trained on three source corpora, PRISM matches or outperforms REVE (92 datasets, 60,000+ hours) on the majority of tasks, demonstrating that targeted diversity can substitute for indiscriminate scale and that dataset count is a confounding variable in model comparison. Second, on a clinically challenging and previously untested task—distinguishing epilepsy from diagnostic mimickers via interictal EEG—the diverse checkpoint out- performs the narrow-source checkpoint by +12.3 p balanced accuracy, the largest gap across all evaluations. Third, systematic inconsistencies between EEG-Bench and EEG-FM-Bench reverse model rankings on identical datasets by up to 24 p; we identify six concrete sources in- cluding split construction, checkpoint selection, segment length, and normalization, showing these factors compound non-additively. Keywords: EEG Foundation Model· Population Heterogeneity· Epilepsy Differential Diagnosis· Self-Supervised Learning· Masked Autoencoder· Benchmark Standardization. Jeet Bandhu Lahiri, Parshva Runwal, and Arvasu Kulkarni contributed equally as first authors. Mahir Jain and Aditya Ray Misra contributed equally as second authors. arXiv:2603.02268v1 [cs.LG] 28 Feb 2026 2Lahiri et al. 1 Introduction EEG foundation models—self-supervised models pretrained on scalp EEG—have produced strong transfer results across motor imagery, sleep staging, and abnor- mality detection [1,2,3,4]. Yet two fundamental questions remain unaddressed. First, what do these representations encode? Pretraining corpora are drawn over- whelmingly from TUH, PhysioNet, and MOABB—repositories predominantly from Europe and North America where recording conditions, amplifier character- istics, and population physiology are confounded within a narrow distribution [4]. A model that achieves low reconstruction error on such data may have memorized recording-environment regularities rather than neural dynamics, and standard benchmarks from the same ecosystem cannot distinguish these explanations. Second, can we trust existing benchmark rankings? EEG-Bench [5] and EEG- FM-Bench [6] are the two most widely used standardized evaluation frameworks for EEG foundation models, yet they make different methodological choices in ways that produce not merely quantitative discrepancies but outright ranking reversals on the same models and datasets. We investigate both questions through controlled ablations with PRISM, a masked autoencoder inspired from the REVE architecture [4]. Three findings emerge. First, population heterogeneity does not uniformly improve downstream accuracy—it changes what representations are good for. We construct a narrow- source pool (TUH + PhysioNet) and a geographically diverse pool adding multi- center South Asian clinical recordings. Under linear probing, the narrow-source checkpoint leads on distribution-matched benchmarks; under fine-tuning, the diverse checkpoint matches or exceeds it on most tasks. Notably, PRISM trained on three source corpora matches or outperforms REVE trained on 92 datasets, demonstrating that targeted data diversity can substitute for indiscriminate data scale—and that dataset count is a confounding variable in EEG foundation model comparison. Second, we provide a concrete analysis of six specific methodological differ- ences between EEG-Bench and EEG-FM-Bench that produce divergent results: train/validation split construction, checkpoint selection strategy, input segment length, preprocessing normalization, classification head selection, and self-reported versus standardized evaluation. We show that correcting individual factors re- solves discrepancies for some models but not others, revealing compounding interactions. These inconsistencies underscore the urgent need for a community consensus on evaluation methodology—without which model comparisons are unreliable. Third, we introduce the first evaluation of an EEG foundation model on distinguishing epilepsy from its diagnostic mimickers—including psychogenic non-epileptic seizures (PNES), syncope, and other paroxysmal disorders—using routine interictal EEG, a clinically relevant task that has not previously been addressed by any foundation model. This task is substantially harder than ex- isting benchmarks: up to 25% of patients with refractory epilepsy diagnoses are later found to have been misdiagnosed [7], and the mean diagnostic delay for Title Suppressed Due to Excessive Length3 PNES alone exceeds five years [8]. On this task, the diverse checkpoint signifi- cantly outperforms the narrow-source checkpoint, providing direct evidence that pretraining population composition impacts clinically meaningful downstream performance. 2 Related Work EEG foundation models. Self-supervised pretraining for EEG has advanced through BIOT [1] (cross-data transformer on TUH), LaBraM [2] (vector-quantized codebook across TUH and MOABB), CBraMod [3] (criss-cross attention), and REVE [4] (92 datasets, 25,000 subjects, 4D Fourier positional encoding enabling transfer across arbitrary montages—an architectural innovation we adopt). All share a limitation: pretraining data originates predominantly from EU/US repos- itories, and no prior work has tested whether broadening the population changes what the model learns. Pretraining data scale as a confounding variable. A prevailing assump- tion in EEG foundation model development is that larger pretraining corpora yield better downstream representations, mirroring scaling intuitions from lan- guage modeling. This has driven a progression from single-corpus models trained on TUH alone to REVE’s corpus spanning 92 datasets, 25,000 subjects, and over 60,000 hours. However, dataset quantity and representation quality are not equiv- alent, and the EEG domain has not validated that scaling laws from language transfer to neural signals. EEG-FM-Bench itself observes that “scaling data often fails to yield proportional gains in downstream tasks” [6], and empirical work on EEG pathology classification has found ceiling effects when scaling labeled data beyond moderate corpus sizes [29]. Critically, when models differ simultaneously in architecture, pretraining objective, corpus size, and corpus composition, it is impossible to attribute performance differences to any single factor. No prior work has performed controlled dataset ablations—varying corpus composition while holding architecture, objective, and evaluation protocol fixed—to establish what data actually contributes. This confound renders published dataset-count comparisons between models largely uninterpretable as evidence of model quality. Evaluation inconsistencies. EEG-Bench [5] and EEG-FM-Bench [6] each provide standardized evaluation protocols but make different choices in prepro- cessing, segmentation, checkpoint selection, and classifier heads. As we show in Section 4.5, these choices can reverse model rankings on the same dataset, with discrepancies exceeding 24 percentage points. This mirrors findings in medical imaging, where evaluation methodology determines whether a model appears to generalize [15]. We provide the first systematic decomposition of the specific sources of these inconsistencies and their interactions. Demographic bias in medical AI. Chest X-ray models exhibit systematic underdiagnosis of marginalized subgroups [13], and most clinical AI algorithms train on data from only three geographic regions [14]. These failures arise from dis- tributional shortcuts—models encode hospital site rather than pathology [15,16]. The analogous concern for EEG has not been tested until this work. 4Lahiri et al. Epilepsy vs. mimickers. Distinguishing epilepsy from its mimickers is among the hardest problems in clinical neurophysiology. PNES accounts for 20– 30% of epilepsy monitoring unit admissions [9], with diagnostic delays exceeding five years [8] and misdiagnosis rates of up to 25% [7]. While prior ML work addresses seizure detection [12], no study has applied EEG foundation models to differential diagnosis from interictal recordings. 3 Method Our experimental design isolates the effect of pretraining population on down- stream representation quality, evaluating across both standard benchmarks and a novel clinical task. 3.1 Pretraining Pools We construct two dataset pools with identical preprocessing (200 Hz resampling, 0.5–99.5 Hz bandpass, notch filtering at 50 and 100 Hz, per-channel Z-score normalization, ±15σ clipping): D1 (Narrow-source): TUH Corpus [18] + PhysioNet Motor Imagery [19]— the standard pretraining distribution used by most EEG foundation models, predominantly from North American and European recording centers under uniform clinical protocols. D2 (Multi-source): D1 augmented with multi-center clinical recordings from 9663 subjects totalling 4170 hours from South Asian institutions, introducing geographic, demographic, and acquisition-system heterogeneity—different EEG systems (including Natus, Nihon Kohden, and BPL units), referencing conventions, and electrode impedance profiles. No balancing or re-weighting is applied. All recordings are mapped to the standard 10–20 montage (19 channels) via automated channel-name normalization handling common aliases (T7→T3, P7→T5) and acquisition-specific prefixes. 3.2 Architecture PRISM adopts the masked autoencoder architecture introduced by REVE’s 4D positional encoding [4], with modifications to the auxiliary training path. Each EEG channelcwith signal x c ∈R T is segmented into patches ofP=200 samples (1 s at 200 Hz) with overlapO=20 samples (step sizeS=P−O=180), and each patch p∈R P is embedded via a convolutional projection e = W e p where W e ∈R D×P with D=512. 4D Positional Encoding. Following REVE [4], each token receives a po- sitional encoding from its electrode’s spatial coordinates (x,y,z) and temporal patch indext. Let c = [x,y,z,t] ⊤ ∈R 4 denote the coordinate vector. We con- struct a frequency matrix F∈R 4×K from the Cartesian product ofn freq =4 linearly spaced frequencies per dimension, yielding K=n 4 freq =256 combinations: PE(c) = LN W f sin(F ⊤ c); cos(F ⊤ c) + MLP(c) (1) Title Suppressed Due to Excessive Length5 where W f ∈R D×2K is a learned projection and LN is layer normalization. This encoding enables processing of arbitrary electrode montages without retraining. Encoder–Decoder. The encoder is a 12-layer pre-norm transformer (8 heads, D=512, GELU feedforward with expansion factor 4) processing only unmasked tokens. The decoder (4 layers, same configuration) receives encoder outputs at visible positions and learnable mask tokens at masked positions, reconstructing masked patches via a linear head. Masking. Spatio-temporal block masking at ratioρ=0.55 selects random seed tokens and masks all tokens within a 3 cm spatial and 3 s temporal radius, iterating until⌊ρN⌋tokens are masked; excess masked tokens are randomly restored to enforce exact batch-uniform counts. 3.3 Training Objective The model is trained with two L1 reconstruction losses. Let ˆ x i and x i denote the reconstructed and original signal for masked patchi, and letMdenote the set of masked indices. The primary loss measures decoder reconstruction quality: L pri = 1 |M| X i∈M ∥ ˆ x i − x i ∥ 1 (2) An auxiliary path concatenates feedforward outputs from all encoder layers per token and pools them via a learned attention mechanism with a single query vector to produce a global embedding, which independently reconstructs each masked patch conditioned on the masked position’s positional encoding. The auxiliary lossL sec follows the same L1 form; by forcing reconstruction from a single global token, this path distributes information across encoder depth. The total loss is: L =L pri + λL sec , λ = 0.1(3) Two checkpoints are trained: D1 (narrow-source) and D2 (multi-source), both with identical architecture and training hyperparameters. 3.4 Downstream Tasks and Evaluation Benchmark tasks. We evaluate on six established tasks: ADFTD [26] (Alzheimer’s/ frontotemporal dementia, 2-class), BCI-IV-2a [20] (motor imagery, 4-class), HMC [21] (sleep staging, 5-class), PhysioNet-MI [19] (motor imagery, 4-class), Siena Scalp [27] (pathology, 2-class), and EEGMAT [22] (cognitive load, 2-class). Clinical task: Epilepsy vs. Mimickers. We introduce a novel downstream task addressing a critical unmet clinical need. Distinguishing epilepsy from its diagnostic mimickers—primarily psychogenic non-epileptic seizures (PNES), vasovagal syncope, and other paroxysmal non-epileptic events—from routine interictal EEG represents one of the most diagnostically challenging problems in clinical neurophysiology. Unlike seizure detection tasks that operate on ictal recordings (which contain overt electrographic correlates), interictal EEG in this 6Lahiri et al. EEG Raw Signalx (B, C, T) Patch Embed \+ Transformer Encoder 12 Layers, 8 Heads (Visible Tokens Only) /+ MAE Decoder 4 Layers, 8 Heads Linear Predictor Primary Loss MainL 1 Recon Loss 3D Topoplot Positions (B, C, 3) 4D PosEnc Fourier + MLP Block Masking Ratio: 0.55 Random Seed Boolean Mask Keeps Visible Learnable [QUERY] Attention Pooling Intermediate Layer Outputs + Aux Predictor Global Emb + PE Secondary Loss AuxL 1 Recon Loss Total Loss L main + 0.1L aux Legend \ Split into masked/visible / Join visible/masked tokens + Element-wise addition Positional Encoding Path Mask Indices Path Fig. 1. PRISM architecture. Raw EEG is patchified per channel and embedded. 4D positional encoding [4] (electrode coordinates + temporal index) is added before the encoder. The decoder fills masked positions with learnable tokens and reconstructs signal. An auxiliary path pools intermediate encoder outputs via learned attention, producing a global embedding that independently reconstructs masked patches. differential diagnosis problem is subtle: background activity may appear broadly similar across groups, and discriminative markers are sparse, non-specific, and highly dependent on careful reading. This setting has received no prior attention from EEG foundation model research, making it an important and underexplored benchmark for clinically meaningful evaluation. The dataset comprises 200 subjects—100 with confirmed epilepsy and 100 with confirmed mimicker diagnoses—recorded on clinical-grade Natus EEG systems at 256 Hz in the standard 10–20 montage at tertiary care clinical centers in India. All subjects are from the South Asian population, making this dataset geographically and demographically distinct from existing EEG benchmarks. Diagnosis was established by board-certified epileptologists following established criteria [25], using the full clinical workup including video-EEG monitoring, ictal semiology assessment, longitudinal follow-up, and response to treatment. Cases where diagnosis remained uncertain at follow-up were excluded, ensuring label reliability. We use subject-level splits: 85 subjects per class for training, 5 for validation, and 10 for held-out evaluation. Data is resampled to 200 Hz and preprocessed identically to the pretraining pipeline. The dataset will be released as a public benchmark upon acceptance, constituting the first publicly available interictal EEG dataset curated for this differential diagnosis task; details regarding data access will be made available via the project repository at publication. Adaptation strategies. We evaluate four regimes. Linear probing (LP): encoder frozen, only the classification head is trained. Full fine-tuning, single-stage (Full-Single): all parameters updated end-to-end from the pretrained checkpoint. Full fine-tuning, dual-stage (Full-Dual): first train the classification head with frozen encoder (LP), then unfreeze all parameters for joint training—preserving pretrained structure during initial adaptation. Partial fine-tuning, single-stage Title Suppressed Due to Excessive Length7 Table 1. Effect of pretraining population on balanced accuracy across six benchmark tasks. LP = linear probing (frozen encoder, MLP head), FT = full fine-tuning single- stage with MLP head. Bold = best, underline= second best per section-column. Ckpt ADFTD BCI-IV-2a HMC PhysioNet Siena EEGMAT LP D1 0.554 0.4810.682 0.604 0.650 0.576 D2 0.5330.507 0.6790.5930.6380.649 FT D1 0.5240.5420.703 0.6310.738 0.592 D2 0.545 0.543 0.6730.640 0.738 0.648 Table 2. Head ablation on D2, full fine-tuning. Bold = best,underline= second best per column. HeadADFTD BCI-IV-2a HMC PhysioNet Siena EEGMAT Attn Pool 0.4850.4100.707 0.604 0.7370.636 Avg Pool 0.5270.389 0.7050.6270.787 0.607 MLP0.545 0.543 0.673 0.640 0.738 0.648 (Partial-Single): only the finalkencoder layers and classification head are updated, keeping early layers frozen. Classification head ablation is performed across attention pooling, average pooling, and a two-layer MLP. All results report balanced accuracy. 4 Experiments and Results 4.1 Study 1: Effect of Population Composition Table 1 reveals a dissociation invisible when a single evaluation regime is reported. Under LP, D1 outperforms D2 on four of six tasks, consistent with D1’s represen- tations being aligned with evaluation data statistics. Under FT, D2 matches or exceeds D1 on five of six tasks. Diverse pretraining produces representations in a more general embedding region that requires nonlinear adaptation but ultimately yields superior performance. This is consistent with the evaluation instability doc- umented by EEG-Bench and EEG-FM-Bench, where different pooling strategies and fine-tuning regimes routinely reverse model rankings [5,6]. The exception—HMC sleep staging, where D1 retains its FT advantage—is consistent with the neurophysiological distinctiveness of sleep macrostructure, defined by gross spectral features (delta dominance in N3, spindles in N2 [23]) largely invariant to population and recording system. 4.2 Study 2: Classification Head Ablation Table 2 shows that the MLP head outperforms both pooling alternatives on four of six tasks within our model, with the largest margin on BCI-IV-2a (+13.3 p over attention pooling, +15.4 p over average pooling). The exceptions—HMC 8Lahiri et al. Table 3. Adaptation strategy comparison on D2, MLP head. Full-Single = all parameters trained end-to-end; Full-Dual = LP stage then full fine-tuning; Partial-Single = only last encoder layer unfrozen. Bold = best, underline= second best per column. StrategyADFTD BCI-IV-2a HMC PhysioNet Siena EEGMAT LP0.5330.507 0.679 0.593 0.638 0.649 Full-Dual0.557 0.5210.696 0.598 0.613 0.635 Full-Single0.5450.543 0.673 0.640 0.738 0.648 Partial-Single 0.5310.518 0.6820.6070.6630.658 Table 4. Epilepsy vs. mimickers: balanced accuracy on held-out subjects. D1 = narrow- source checkpoint, D2 = multi-source checkpoint. Full-Single fine-tuning with MLP head. Bold = best, underline= second best. D1 D2 Full-Single 0.4840.607 and Siena—are tasks where class boundaries align with coarse global features that simple aggregation captures adequately. We hypothesize this reflects an interaction between the reconstruction pre- training objective and the aggregation mechanism. MAE tokens are optimized for patch-level reconstruction, encoding local waveform morphology rather than globally discriminative features. Attention pooling may inherit biases toward reconstruction-informative tokens rather than discriminative ones, while average pooling dilutes sparse discriminative content. The MLP applies a learned nonlin- ear re-projection from reconstruction space toward classification. EEG-FM-Bench reports analogous sensitivity to head choice across models [6], suggesting the effect is not unique to our architecture. 4.3 Study 3: Fine-Tuning Strategy No strategy dominates (Table 3). Full-Dual wins on ADFTD and HMC—tasks requiring distributed temporal characterization—where the two-stage approach preserves pretrained structure. Full-Single excels on motor imagery and Siena, where end-to-end feature reorganization around localized discriminative patterns is effective. Partial-Single achieves the best result on EEGMAT, balancing adap- tation with preservation of early feature extraction layers. Adaptation strategy is a consequential experimental choice, not merely a hyperparameter. 4.4 Study 4: Epilepsy vs. Mimickers Table 4 presents results on the most diagnostically challenging task in our evaluation suite. The multi-source checkpoint (D2) outperforms the narrow- source checkpoint (D1) by 12.3 p in balanced accuracy—substantially larger than the gap on any benchmark task (Table 1), where D1 and D2 are typically separated by fewer than 2 p under FT. Title Suppressed Due to Excessive Length9 Table 5. Comparison with EEG foundation models under two standardized evaluation protocols on motor imagery tasks. Protocols differ in preprocessing, segmentation, check- point selection, and evaluation heads—see Section 4.5 for a systematic decomposition. Bold = best, underline= second best per column. BCI-IV-2aPhysioNet-MI Model EEG-Bench FM-Bench EEG-Bench FM-Bench LaBraM0.280.380.250.49 CBraMod 0.310.350.270.46 REVE0.41 0.330.330.52 PRISM0.440.510.320.59 Epilepsy vs. mimicker discrimination from interictal EEG is a task that even expert neurologists find difficult—routine EEG sensitivity for epilepsy is∼50% in isolation [11], and EEG overinterpretation is a well-documented source of misdiagnosis [10]. That pretraining population produces a measurable 12.3 p difference on this task, when the evaluation data itself comes from a single center with uniform protocols, suggests the multi-source checkpoint has learned more robust neural representations that better capture subtle pathological distinctions. We hypothesize that multi-source pretraining forces the encoder to disentangle neural content from acquisition artifacts during reconstruction—an implicit invariance that the narrow-source encoder lacks. Notably, the evaluation cohort is drawn exclusively from South Asian clinical centers, demographic and geographic territory well-represented in D2 but absent in D1. This demographic alignment between pretraining diversity and evaluation population may contribute to D2’s advantage. Future work should disentangle the relative contributions of geographic representation and acquisition-system diversity. 4.5Study 5: Comparison with Existing Models and Cross-Protocol Inconsistency PRISM achieves the highest accuracy on three of four protocol–task combinations (Table 5). However, the cross-protocol inconsistencies are at least as informative as the absolute rankings. Decomposition of protocol differences. EEG-Bench and EEG-FM-Bench differ in six concrete methodological dimensions, each of which can independently shift results: (1) Train/validation split construction. EEG-Bench uses a subject-level held-out test set, with the remaining data split randomly at the segment level (20% validation), meaning some subjects’ segments span both train and validation sets. EEG-FM-Bench maintains strict subject-level separation for all splits. This discrepancy means EEG-Bench’s effective training set is equivalent to EEG-FM- Bench’s combined train and validation sets, inflating EEG-Bench performance estimates for subject-independent generalization. 10Lahiri et al. (2) Checkpoint selection policy. EEG-Bench selects the checkpoint achiev- ing best validation performance (early stopping), while EEG-FM-Bench returns the last training checkpoint. For models prone to overfitting on small fine-tuning sets, this difference can produce large performance swings independent of model quality. (3) Input segment length. EEG-FM-Bench processes 4-second windows whereas EEG-Bench uses 3-second windows by default. The additional second provides more temporal context at inference, systematically favoring models that capture longer-range temporal dependencies—a bias that interacts nontrivially with architectures that use windowed attention. (4) Preprocessing normalization. The two frameworks apply different normalization schemes in their feature extraction pipelines. Normalization differ- ences affect the effective amplitude distribution presented to the model, which can disproportionately affect models that were pretrained under specific normalization assumptions. (5) Classification head and fine-tuning methodology. Self-reported metrics in individual model papers are often produced under the protocol most favorable to that model (e.g., attention pooling rather than MLP heads; specific learning rate schedules). EEG-FM-Bench imposes a fixed evaluation protocol across models, which is more comparable but may disadvantage models whose optimal heads were not included. Our Study 2 (Section 4.2) demonstrates that head choice alone can shift performance by over 15 p within the same model on the same task. (6) Compound interactions. We find that correcting for the validation split and segment length equalizes results for CBraMod on BCI-IV-2a, but analogous corrections do not resolve discrepancies for other models. This indicates that the six factors compound non-additively: a model optimized to exploit one protocol’s structure may be systematically disadvantaged by another’s, making it impossible to attribute ranking differences to any single methodological choice. The comparison with REVE is instructive: both share similar architectural components but differ in pretraining composition and scale. REVE uses 92 datasets and 25,000 subjects from predominantly EU/US sources; PRISM uses three source corpora. On BCI-IV-2a, PRISM outperforms REVE under both protocols. On PhysioNet-MI, rankings reverse between protocols (REVE +1 p EEG-Bench; PRISM +7 p FM-Bench), illustrating that multi-source pretraining provides complementary value that is protocol-dependent in its expression. Dataset quantity as a confound. Beyond protocol inconsistency, this comparison surfaces a structural problem in EEG foundation model evaluation: dataset count is routinely used as a proxy for model quality, yet it is a confounding variable whenever models also differ in architecture, pretraining objective, and corpus composition—as all published model pairs do. PRISM matches or exceeds REVE on three of four protocol–task combinations despite a roughly 30-fold difference in the number of source datasets. This does not imply that data quantity is unimportant; it implies that the marginal return of aggregating additional EU/US datasets past some saturation point may be limited relative to the return Title Suppressed Due to Excessive Length11 of targeted diversity. Without controlled dataset ablations—varying corpus size while holding all other factors fixed—one cannot determine whether REVE’s broader corpus or its other design choices drive its performance on tasks where it leads. The field therefore risks two compounding errors: (i) misattributing performance to data scale when the driving factor may be architecture or objective, and (i) over-collecting data from redundant distributions that do not improve downstream representations. Systematic dataset ablations, in the spirit of our D1-vs-D2 comparison, should be a required component of any EEG foundation model publication. These findings collectively underscore that how a model is evaluated and what data it is trained on can each matter as much as the model itself, and that current reporting practices make it difficult to disentangle these factors [5,6]. 5 Discussion The distribution-matching trap. LP accuracy on narrow-source benchmarks systematically favors narrow-source pretrained models because their represen- tations are pre-aligned with the evaluation distribution. The reversal under FT shows the diverse model encodes equivalent or superior information requiring adaptation to extract. For clinical systems that always undergo task-specific adaptation, LP-only evaluation is misleading. From benchmarks to clinical impact. Pretraining population composition produces its largest effect on the clinically hardest task—epilepsy vs. mimickers. On standard benchmarks, D1 and D2 are often within 1–2 p under FT; on the clinical task, the gap widens to 12.3 p. Benchmark tasks—dominated by gross physiological distinctions—are insufficient proxies for clinical utility. The community would benefit from incorporating challenging differential diagnosis tasks into evaluation suites, and our released dataset is a step toward this goal. The need for standardized evaluation. Our decomposition in Section 4.5 demonstrates that the six methodological differences between EEG-Bench and EEG-FM-Bench interact in ways that prevent attribution of ranking differences to model properties. We recommend the community converge on: (i) strict subject- level splits throughout; (i) fixed checkpoint selection policies (e.g., best validation or last, but uniformly applied); (i) standardized segment lengths per task; (iv) unified preprocessing including normalization; (v) exhaustive head ablations reported per model rather than cherry-picked. Ideally, an openly maintained benchmark platform would allow continuous community evaluation under a fixed protocol, enabling fair longitudinal comparisons as new models emerge. Implicit invariance through data diversity. Multi-source pretraining spans heterogeneous acquisition systems; reconstruction forces the encoder to disentangle neural content from acquisition artifacts—an implicit form of invari- ance learning that simplifies the pipeline relative to explicit augmentation or multi-stage objectives. Dataset quantity is a confounding variable in model comparison. The EEG foundation model literature has trended sharply toward ever-larger 12Lahiri et al. pretraining corpora, with corpus size frequently cited as evidence of model ambition or expected quality. Our results challenge this framing. PRISM, trained on three source corpora, matches or outperforms REVE—trained on 92 datasets and over 60,000 hours—on three of four protocol–task combinations. This is not a claim that data volume is unimportant: larger corpora may still confer advantages on tasks not evaluated here, particularly those requiring exposure to rare pathological patterns. Rather, it is a claim about interpretability: when two models differ simultaneously in architecture, pretraining objective, corpus size, and corpus composition, observed performance differences cannot be attributed to any single cause. Dataset count has become a confounding variable that inflates reported differences between models while obscuring what actually drives representation quality. The practical consequence is two-fold. First, the community risks over- collection: assembling larger corpora from the same EU/US repositories may yield diminishing returns once basic recording-condition diversity is covered, while the cost of data curation and licensing grows linearly. Our D1-vs-D2 ablation shows that adding 9,663 South Asian clinical subjects—a single diverse source— produces a 12.3 p gain on the hardest clinical task and competitive fine-tuning performance across benchmarks, while our D1 checkpoint already matches models pretrained on far larger EU/US corpora. Second, without pretraining dataset ablations as a first-class experimental contribution, it is impossible to distin- guish “more data” from “better data” from “better architecture” in published comparisons. We therefore recommend that future EEG foundation model papers include controlled dataset ablations alongside architecture and objective ablations, treating corpus composition as an independent experimental variable rather than a fixed condition. Limitations. We use a single model size; scaling effects remain unexplored. The multi-source data confounds geography with acquisition systems, making it impossible to attribute improvements to one factor alone. Our epilepsy–mimicker dataset is drawn from South Asian clinical centers; generalization to other demographic groups requires separate validation. The masking ratio was fixed. While we show the MLP head advantage within our model, generalizability to other reconstruction-based architectures requires further investigation. 6 Conclusion Through controlled ablations with fixed architecture, we found that population heterogeneity reshapes EEG representations in ways invisible to linear probing but exposed by fine-tuning. Most importantly, we demonstrated that this effect matters most where it counts: on a clinically challenging epilepsy vs. mimicker differential diagnosis task that no prior foundation model has addressed, the di- verse checkpoint outperforms the narrow-source baseline by 12.3 p. We provided a systematic decomposition of six concrete methodological differences between EEG-Bench and EEG-FM-Bench that produce ranking reversals of up to 24 p on identical models and datasets, demonstrating that benchmark inconsistency Title Suppressed Due to Excessive Length13 is not merely quantitative but qualitative. We further showed that PRISM, trained on three source corpora, matches or outperforms REVE—trained on 92 datasets—on the majority of evaluated protocol–task combinations, demonstrat- ing that dataset count is a confounding variable in model comparison and that targeted diversity can substitute for indiscriminate scale. Our model matches or outperforms larger-scale models on standard benchmarks, and the cross-protocol instability we document reinforces the urgent need for community consensus on evaluation methodology, including dataset ablations as a required experimental component. We will release the 200-subject epilepsy–mimicker dataset to catalyze progress on clinically grounded EEG foundation model evaluation. Model code and implementation are available at:https://osf.io/gv4p3/overview?view _only=363cb103cff44bfc9ce916240ed403a6. References 1. Yang, C., Westover, M., Sun, J.: BIOT: Biosignal transformer for cross-data learning in the wild. In: NeurIPS (2024) 2. Jiang, W., Zhao, L., Lu, B.: Large brain model for learning generic representations with tremendous EEG data in BCI. In: ICLR (2024) 3.Wang, J., et al.: CBraMod: A criss-cross brain foundation model for EEG decoding. arXiv:2412.07236 (2024) 4.El Ouahidi, Y., et al.: REVE: A foundation model for EEG with large-scale pretraining on 25,000 subjects. In: NeurIPS (2025) 5.Kastrati, A., et al.: EEG-Bench: A benchmark for EEG foundation models in clinical applications. arXiv:2512.08959 (2025) 6.Xiong, W., Li, J., Li, J., Zhu, K., Jiang, C.: EEG-FM-Bench: A comprehensive bench- mark for the systematic evaluation of EEG foundation models. arXiv:2508.17742 (2025) 7. Benbadis, S.R.: The differential diagnosis of epilepsy: A critical review. Epilepsy & Behavior 15(1), 15–21 (2009) 8.Reuber, M., Fernandez, G., Bauer, J., Helmstaedter, C., Elger, C.E.: Diagnostic delay in psychogenic nonepileptic seizures. Neurology 58(3), 493–495 (2002) 9.Benbadis, S.R.: The problem of psychogenic symptoms: Is the psychiatric community in denial? Epilepsy & Behavior 6(1), 9–14 (2005) 10.Benbadis, S.R.: Errors in EEGs and the misdiagnosis of epilepsy: Importance, causes, consequences, and proposed remedies. Epilepsy & Behavior 11(3), 257–262 (2007) 11.Smith, S.J.M.: EEG in the diagnosis, classification, and management of patients with epilepsy. J Neurol Neurosurg Psychiatry 76(suppl 2), i2–i7 (2005) 12.Shoeibi, A., et al.: Epileptic seizures detection using deep learning techniques: A review. Int J Environ Res Public Health 18(11), 5780 (2021) 13. Gichoya, J.W., et al.: Demographic bias of expert-level vision-language foundation models in medical imaging. Science Advances 11(13), eadq0305 (2025) 14.Kaushal, A., Altman, R., Langlotz, C.: Geographic distribution of US cohorts used to train deep learning algorithms. JAMA 324(12), 1212–1213 (2020) 15.Zech, J.R., et al.: Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs. Nature Medicine 24(9), 1347–1353 (2018) 16. DeGrave, A.J., Janizek, J.D., Lee, S.I.: AI for radiographic COVID-19 detection selects shortcuts over signal. Nature Machine Intelligence 3(7), 610–619 (2021) 14Lahiri et al. 17.He, K., et al.: Masked autoencoders are scalable vision learners. In: CVPR, p. 16000– 16009 (2022) 18. Obeid, I., Picone, J.: The Temple University Hospital EEG data corpus. Frontiers in Neuroscience 10, 196 (2016) 19. Goldberger, A.L., et al.: PhysioBank, PhysioToolkit, and PhysioNet. Circulation 101(23), e215–e220 (2000) 20. Tangermann, M., et al.: Review of the BCI Competition IV. Front. Neurosci. 6, 55 (2012) 21. Alvarez-Estevez, D., Rijsman, R.M.: Inter-database validation of a deep learning approach for automatic sleep scoring. PLoS ONE 16(8), e0256111 (2021) 22.Zyma, I., et al.: Electroencephalograms during mental arithmetic task performance. Data 4(1), 14 (2019) 23.Berry, R.B., et al.: The AASM Manual for the Scoring of Sleep and Associated Events: Rules, Terminology and Technical Specifications, Version 2.4. American Academy of Sleep Medicine (2017) 24. Norori, N., et al.: Addressing bias in big data and AI for health care. Patterns 2(10), 100347 (2021) 25.LaFrance, W.C., et al.: Minimum requirements for the diagnosis of psychogenic nonepileptic seizures: A staged approach. Epilepsia 54(11), 2005–2018 (2013) 26. Miltiadous, A., Tzimourta, K.D., Afrantou, T., Ioannidis, P., Grigoriadis, N., Tsa- likakis, D.G., Angelidis, P., Tsipouras, M.G., Glavas, E., Giannakeas, N., et al.: A dataset of scalp EEG recordings of Alzheimer’s disease, frontotemporal dementia and healthy subjects from routine EEG. Data 8(6), 95 (2023) 27.Detti, P.: Siena Scalp EEG Database (Version 1.0.0). PhysioNet (2020).https: //doi.org/10.13026/5d4a-j060 28.Detti, P., Vatti, G., Zabalo Manrique de Lara, G.: EEG synchronization analysis for seizure prediction: A study on data of noninvasive recordings. Processes 8(7), 846 (2020) 29.Kiessner, A.K., Schirrmeister, R.T., Boedecker, J., Ball, T.: Reaching the ceiling? Empirical scaling behaviour for deep EEG pathology classification. Computers in Biology and Medicine 178, 108681 (2024)

Instant research discovery