Separating personal from population gains when calibrating EEG foundation models for new users

Xilin Tao; Kani Chen

cs.LGSep 28, 2026

Separating personal from population gains when calibrating EEG foundation models for new users

Authors: Xilin Tao, Kani Chen

Organizations: Department of Mathematics, The Hong Kong University of Science and Technology, Hong Kong SAR, China · Department of Industrial Engineering and Decision Analytics, The Hong Kong University of Science and Technology, Hong Kong SAR, China

Abstract

Foundation models are increasingly adapted to individual users, but an apparent personalization gain can simply reflect a stronger population model. This distinction matters for brain-computer interfaces, where every new user must be calibrated. We evaluated personal adaptation of three frozen EEG foundation models (CBraMod, REVE and LaBraM) in 235 held-out subjects from three motor-imagery datasets, comparing each subject's adapter with the population model and with adapters fitted to other subjects. Using all first-half session labels, personal adapters improved mean balanced accuracy over the population model by 1.5-5.4 percentage points and outperformed exchanged adapters by 2.3-7.3 points in all nine model-dataset combinations. The size of this benefit depended on population training: with four times the original budget, median gains remained positive (1.0-2.0 points) but were smaller for every model, and no population model reached a confirmed plateau. Acquiring the benefit cheaply was unreliable: few-label calibration was consistently non-negative on only one dataset, and in CBraMod neither unlabeled context nor meta-learned initialization outperformed matched controls. Personalization should therefore be evaluated against both a population reference and exchanged parameters, across population-training budgets.

Figures & tables

Dataset	$N$	Session	EEG / readout	Fit / query trials	Rest
PhysioNet	103	single	64 / 27	22 / 23	EO
Dreyer	80	single	27 / 27	80–120 / 80–120	EO
Cho	52	single	64 / 27	100–120 / 100–120	EO
Lee	54	1	62 / 26	100 / 100	EO
BNCI	9	2	22 / 19	72 / 72	EO

Table 1: Analysis populations and CBraMod input/readout channels. EO denotes eyes-open rest. REVE and LaBraM use native-input pipelines on the core datasets; Lee and BNCI are CBraMod-only extensions. Fit/query counts are trials per subject in the temporal splits, including their ranges. Dataset sources are cited in Methods.

Model	Dataset	Own - population	Own - exchanged
CBraMod	PhysioNet	1.64 [0.77, 2.51]	2.55 [1.74, 3.40]
CBraMod	Dreyer	3.58 [2.94, 4.24]	5.28 [4.64, 5.93]
CBraMod	Cho	5.13 [3.13, 7.46]	6.43 [4.34, 8.83]
REVE	PhysioNet	1.55 [0.91, 2.19]	2.59 [2.02, 3.19]
REVE	Dreyer	5.38 [4.38, 6.43]	7.25 [6.21, 8.33]
REVE	Cho	3.35 [2.26, 4.65]	4.38 [3.25, 5.70]

Table 2: Mean personal contrasts at the initial budgets, in pp, with subject-bootstrap 95% confidence intervals in brackets. Each subject contributes one seed-averaged paired difference. Sample sizes are 103, 80 and 52 for PhysioNet, Dreyer and Cho, respectively.

Cohort	Fold	Train	Validation	Test
Core cohort	0	169	18	48
Core cohort	1	169	18	48
Core cohort	2	170	18	47
Core cohort	3	171	18	46
Core cohort	4	171	18	46
Lee	0	39	4	11

Table 3: Subject-disjoint partitions. The core cohort combines PhysioNet, Dreyer and Cho. Counts are people; seeds do not add independent subjects.

Code	Meaning
Quantities
BA; accuracy	Balanced accuracy; unbalanced classification accuracy. BA is the mean class recall. Absolute scores are percentages; score contrasts use percentage points (pp).
Delta; U	Own-minus-population benefit; own-minus-exchanged personal specificity, respectively, defined in Methods 4.1.
s; d(s); S_s; Q_s; c_s	Target subject; assigned donor; earlier labeled support; later query set; eligible unlabeled context.
P; P_b; b	Named population reference; reference at population-training budget b; original, doubled or quadrupled update budget.
f_theta; phi; a_s; a_d(s)	Frozen backbone; population parameters; target personal parameters; donor personal parameters.

Table S1: Notation and identifier key. Capitalization distinguishes model families, scored conditions and experiment stages.

Model	Dataset	Contrast / condition	N	Median [95% CI]	Mean [95% CI]
CBraMod	PhysioNet	Personal benefit	103	0.98 [0.08, 2.42]	1.64 [0.77, 2.51]
CBraMod	PhysioNet	Personal specificity	103	1.87 [1.51, 2.61]	2.55 [1.74, 3.40]
CBraMod	PhysioNet	Population gain	103	7.77 [6.82, 10.08]	9.14 [7.54, 10.78]
CBraMod	Dreyer	Personal benefit	80	3.29 [2.67, 3.83]	3.58 [2.94, 4.24]
CBraMod	Dreyer	Personal specificity	80	4.89 [4.17, 5.95]	5.28 [4.64, 5.93]
CBraMod	Dreyer	Population gain	80	9.08 [7.33, 10.00]	9.18 [8.23, 10.15]

Table S2: Subject-bootstrap intervals for all main-text comparisons. Population balanced accuracy is in percent; paired differences are in percentage points.

Dataset	Start	Labels	$N$	Mean gain (pp)	Subject SD (pp)
PhysioNet	B3	5	103	0.78	2.98
PhysioNet	B3	10	103	0.93	3.36
PhysioNet	B3	20	103	1.50	3.91
PhysioNet	G	5	103	0.81	3.07
PhysioNet	G	10	103	0.83	3.11
PhysioNet	G	20	103	1.57	3.65

Table S3: CBraMod few-label calibration: paired subject means and between-subject standard deviations. Gains are in percentage points.

Work	Reference	Question	Subject / target setting	Verified control scope
Identity Trap	[ 12 ]	Representation / identity audit	Task- and dataset-dependent	Identity diagnostics; attribution package not established here
EEG-FM-Compass	[ 4 ]	Benchmark and adaptation comparison	LOSO and within-subject few-shot separately	Separate protocols; no joint own/population/swapped gain decomposition
Stacked LoRA	[ 13 ]	Shared and personal low-rank adapters	Subjects seen in pooled trial split	Global adapter comparison; different capacity; mismatch not reported in read text
Lopes et al. (Scientific Reports)	[ 14 ]	Personal encoders and alignment	LOSO; target calibration	Low-rank, shared-adaptation, mismatch, capacity and normalization controls
Nguyen et al.	[ 15 ]	Frozen-model population LoRA	Held-out subjects	Head versus LoRA; no target personal adapter in evaluated pipeline
ResTL	[ 6 ]	Rest-based generative transfer	LOSO; target rest	Noise/rest control; updates decoder; donor-rest swap not reported in read text

Table S4: Related evaluation questions, target settings and controls. The comparison concerns study designs and inspected controls, not an accuracy ranking.

Dimension	Lopes et al., published article [ 14 ]	Present study
Population split	LOSO; validation holds out 20% of trials within each source subject (p. 8, implementation details).	Downstream train/validation/test subjects disjoint; core pools three datasets.
Target labels	First 48 trials for encoder selection and fine-tuning (p. 7, Eq. 10; Fig. 2).	Full earlier half and specified few-label prefixes; later half for evaluation.
Parameters	Subject-specific encoder branches and residual input-channel low-rank layers (pp. 6–8; Table 1, p. 13).	Frozen pretrained transformers with personal attention LoRA; CBraMod FiLM extension.
Shared reference	Shared encoder (p. 6); the discussion reports no encoder-bank accuracy advantage over a fine-tuned shared baseline under matched adaptation (p. 20).	Named population reference and ordinary-continuation sensitivity, with paired BA contrasts.
Mismatch	Class distinctiveness and matched/mismatched latent geometry; source-subject routing (pp. 14–16, Figs. 6B and 8).	Held-out target query BA using owner versus donor adapters fitted on earlier support labels.
Capacity	Widened shared model and subject-specific batch-normalization controls (p. 15, after the matched/mismatched analysis).	Same-capacity true/shuffled/default context; population and personal fits have different capacity.

Table S5: Control-by-control comparison with the published Lopes et al. study. Page numbers refer to the publisher PDF described above.

ID	Check	Preserved outcome
E01	Split, temporal, shuffle, identity and statistics contracts	Implementation checks passed; synthetic checks do not establish task performance.
E02	Checkpoint source and compute compatibility	Execution checks passed; pretraining overlap remains a separate source audit.
E03	Initial data-screening stop rule	Triggered for PhysioNet; the initial implementation also rejected Cho. Screening was revised before the subsequent analyses; the initial failures remain reported.
E04	Initial task-head accuracy stop rule	Saved head results exceed the threshold; the later record-only rule does not erase the previous data-screening failure.
E05	Matched eyes-closed versus eyes-open alpha diagnostic	Most PhysioNet records agree; exceptions retained. Dreyer/Cho do not support the same diagnostic.
E06	Rest availability, state, boundaries, duration and finite values	Included subjects and exceptional boundaries preserved; missing records are not marked as passed.

Table S6: Engineering and data checks with their preserved outcomes.

Gate ID	Cohort	Prespecified rule	Observed	Decision
B-G	pooled	Median vs B0 >= 2 pp	median 1.000	Fail
B2-film_first2	pooled	Median vs B0 >= 2 pp; one-sided Holm(4) vs tuned head < .05	median 1.833; pH 0.030713	Fail
B2-film_all	pooled	Median vs B0 >= 2 pp; one-sided Holm(4) vs tuned head < .05	median 2.833; pH 1.155e-08	Pass
B2-lora4	pooled	Median vs B0 >= 2 pp; one-sided Holm(4) vs tuned head < .05	median 2.800; pH 1.2765e-15	Pass
B2-lora8	pooled	Median vs B0 >= 2 pp; one-sided Holm(4) vs tuned head < .05	median 3.409; pH 1.3182e-17	Pass
C-M_film	pooled	One-sided Holm(4) vs B4 and B0 < .05; recovery >= .30; drop > 2 pp <= .10	both p pass False; recovery 0.981; drop 0.230	Fail

Table S7: Original joint decision rules, observed components and decisions. A favorable component does not override a failed joint rule.

ID	Item	Status and interpretation
N01	Initial success reference B1	Replaced by B0 before the subsequent experiment; success was not assessed retrospectively against B1.
N02	Meta-learning confirmation on Lee	Not executed because the M primary gate failed. The separate W Lee evaluation is not that confirmation.
N03	Cross-session robustness	Not executed; current performance evidence is within-session.
N04	Context-duration curve	Planned mechanism stage not started; no result or negative test inferred.
N05	Eyes-open, eyes-closed and combined context	Not executed under the matched-duration design.
N06	Rest-versus-task contextual generator	Not executed; task-neighbor transfer is a different operation.

Table S8: Unexecuted, unavailable and superseded items with their interpretation.

Cohort	Family	Seed	Run steps	Selected	Train flat	Val. flat	Both flat
Core cohort	FiLM population	11	2750	750	False	False	False
Core cohort	LoRA population	11	2000	0	False	False	False
Core cohort	FiLM population	23	2250	250	False	False	False
Core cohort	LoRA population	23	2500	500	False	False	False
Core cohort	FiLM population	37	2500	500	False	False	False
Core cohort	LoRA population	37	2000	0	False	False	False

Table S9: CBraMod continuation records by dataset, parameter family, fold and seed. Update counts are additional continuation updates; no run confirms the joint plateau.

Method	Inner steps	Outer updates	Mean seconds	Runs
M1	5	2000	704.2	25
M1	10	2000	1250.1	25
M1	20	2000	2342.3	25
M2	5	2000	695.4	25
M2	10	2000	1235.3	25
M2	20	2000	2308.5	25

Table S10: Recorded CBraMod training times. M1 and M2 include inner-loop work despite matching R2 in scheduled outer-update budget. Times are per-run means, not total study costs.

Component	Setting (semicolon order)	Value / grid
Partitions	Folds; seed values	5; 11, 23, 37, 53, 71
Task head	Learning rates; epochs; batch	0.001, 0.0003; 50; 128
Task head	Weight decay; MLP hidden width	0.01; 128
CBraMod input	Hz; high-pass Hz; rest segment s	200; 0.3; 4
Mains notch	PhysioNet; Dreyer; Cho (Hz)	60; 50; 60
Personal diagnostics	Earlier FiLM variants	film2, film4, beta4, head

Table S11: Recorded architecture and optimization settings. Semicolons in each value cell follow the setting-column order; commas enumerate a search grid. Forty-label conditions require sufficient earlier support. New-model entries refer to REVE and LaBraM; personal optimization follows the common grid.

Dataset	Model	Source	G-B0 mean	Own-G mean	Own-swap mean	Own-swap median
PhysioNet	CBraMod	unlisted	+9.14	+1.64	+2.55	+1.87
PhysioNet	REVE	unlisted	+13.92	+1.55	+2.59	+2.39
PhysioNet	LaBraM	listed	+20.21	+1.52	+2.35	+2.09
Dreyer	CBraMod	unlisted	+9.18	+3.58	+5.28	+4.89
Dreyer	REVE	listed	+8.98	+5.38	+7.25	+6.78
Dreyer	LaBraM	unlisted	+9.31	+3.91	+5.41	+4.32

Table S12: Mean and median core contrasts and pretraining-source-list status. Differences are in percentage points; seeds are averaged within subject.

Model	N	Median U (pp)	Original Holm p	Decision
REVE	235	3.50	1.38e-35	Pass
LaBraM	235	3.17	1.19e-29	Pass

Table S13: Original pooled specificity decisions for REVE and LaBraM. The saved three-slot correction is retained.

Model	Dataset	Contrast	N	Mean	Median	SD	Raw p
REVE	Cho	G minus B0	52	2.479	1.633	4.207	6.83e-05
REVE	Cho	own minus G	52	3.353	2.600	4.472	1.27e-08
REVE	Cho	own minus swap	52	4.383	3.480	4.556	2.57e-10
REVE	Dreyer	G minus B0	80	8.978	8.833	8.121	5.54e-12
REVE	Dreyer	own minus G	80	5.381	4.667	4.725	6.75e-14
REVE	Dreyer	own minus swap	80	7.251	6.775	4.911	4.39e-15

Table S14: REVE and LaBraM core contrasts by dataset. Means, medians and standard deviations are in percentage points; p-values are the saved descriptive raw values.

Dataset	Model	Descriptive stability
PhysioNet	CBraMod	Stable
PhysioNet	REVE	Stable
PhysioNet	LaBraM	Stable
Dreyer	CBraMod	Not stable
Dreyer	REVE	Stable
Dreyer	LaBraM	Not stable

Table S15: Descriptive consistency of few-label calibration across the prescribed label budgets.

Dataset	Labels	N	Mean	Median	SD	Decline fraction	Raw p
Cho	5	52	-0.383	-0.400	2.036	0.154	0.963
Cho	10	52	0.241	0.000	2.574	0.115	0.409
Cho	20	52	0.928	0.700	2.212	0.077	0.00296
Cho	40	52	1.247	0.733	2.855	0.038	0.00164
Dreyer	5	80	0.544	0.333	3.523	0.138	0.0253
Dreyer	10	80	1.363	0.917	3.418	0.087	0.000253

Table S16: REVE few-label results. Gains are in percentage points; decline is the subject fraction losing more than two points.

Dataset	Labels	N	Mean	Median	SD	Decline fraction	Raw p
Cho	5	52	0.680	0.400	4.006	0.154	0.182
Cho	10	52	1.860	0.900	5.147	0.135	0.00777
Cho	20	52	2.956	1.700	5.768	0.077	4.08e-05
Cho	40	52	3.194	1.717	6.072	0.077	1.19e-05
Dreyer	5	80	-0.175	0.000	2.922	0.175	0.518
Dreyer	10	80	-0.141	0.000	2.769	0.163	0.4

Table S17: LaBraM few-label results. Gains are in percentage points; decline is the subject fraction losing more than two points.

Family	Current evidence and status
Source and structure audits	Completed source and input/adapter audits; source overlap remains visible. Lee full-channel extension remains unexecuted.
Engineering and formal entry	Saved loading, freezing, selection, prediction, and source checks completed; engineering checks alone are not performance evidence.
Resource accounting	Completed saved allocation audit includes unsuccessful cache work; shared multi-dataset training costs are counted once.
G trajectories	All completed new-model trajectories fail to establish the specified plateau; selected and run steps are listed below.
Core specificity	Both new models pass the original pooled gate; own-minus-G is separately reported.
Core population and personal gains	Every model/dataset core comparison is present above; dataset-level tests remain descriptive.

Table S18: Cross-model scope of the specified evaluation families.

Model	Fold	Seed	Initial	Selected updates	Extra run	Stop	Plateau
REVE	0	11	3000	1500	3500	Loss patience	No
REVE	0	23	3000	1500	3500	Loss patience	No
REVE	0	37	3000	250	2250	Loss patience	No
REVE	0	53	1000	1000	3000	Loss patience	No
REVE	0	71	3000	0	2000	Loss patience	No
REVE	1	11	3000	1250	3250	Loss patience	No

Table S19: REVE and LaBraM continuation records. Selected and executed updates are distinguished; the stopping rule does not establish a joint plateau.

Dataset	Budget	N	G mean (SD)	Delta mean (SD)	Median	U mean (SD)	Median
PhysioNet	1x	103	80.99 (11.70)	1.63 (4.59)	0.985	2.55 (4.35)	1.871
PhysioNet	2x	103	81.02 (11.88)	0.94 (5.15)	0.758	2.26 (4.53)	1.932
PhysioNet	4x	103	80.98 (11.57)	0.76 (4.55)	0.227	2.03 (3.91)	1.742
Dreyer	1x	80	80.37 (8.46)	3.58 (2.99)	3.292	5.28 (2.94)	4.892
Dreyer	2x	80	82.18 (8.30)	2.93 (2.93)	2.500	4.71 (3.03)	4.442
Dreyer	4x	80	82.21 (8.42)	3.09 (3.10)	2.333	4.33 (2.95)	3.975

Table S20: Completed CBraMod population-budget results. Population balanced accuracy is in percent; personal benefit and specificity are in percentage points.

Dataset	Budget	N	G mean (SD)	Delta mean (SD)	Median	U mean (SD)	Median
PhysioNet	1x	103	84.71 (11.20)	1.55 (3.32)	0.985	2.59 (3.05)	2.386
PhysioNet	2x	103	84.87 (10.79)	1.19 (3.59)	0.769	2.26 (3.33)	1.769
PhysioNet	4x	103	82.56 (10.97)	0.18 (2.80)	0.000	1.42 (3.08)	1.288
Dreyer	1x	80	83.91 (8.37)	5.38 (4.73)	4.667	7.25 (4.91)	6.775
Dreyer	2x	80	85.75 (7.65)	4.09 (3.93)	3.583	6.51 (4.56)	5.642
Dreyer	4x	80	82.70 (8.34)	3.16 (3.71)	1.750	4.74 (4.02)	3.792

Table S21: Completed REVE population-budget results. Population balanced accuracy is in percent; personal benefit and specificity are in percentage points.

Dataset	Budget	N	G mean (SD)	Delta mean (SD)	Median	U mean (SD)	Median
PhysioNet	1x	103	82.29 (12.04)	1.52 (3.63)	0.909	2.35 (3.79)	2.091
PhysioNet	2x	103	82.78 (11.89)	1.10 (3.85)	0.909	2.27 (3.81)	2.030
PhysioNet	4x	103	82.48 (12.18)	1.21 (4.48)	0.303	2.53 (4.41)	2.129
Dreyer	1x	80	83.17 (8.06)	3.91 (3.79)	3.000	5.41 (3.70)	4.317
Dreyer	2x	80	83.47 (8.42)	3.94 (3.68)	3.417	4.35 (3.30)	3.467
Dreyer	4x	80	84.34 (7.91)	3.30 (3.15)	2.543	4.05 (3.00)	3.283

Table S22: Completed LaBraM population-budget results. Population balanced accuracy is in percent; personal benefit and specificity are in percentage points.

Model	Budget	Six-slot Holm p	Curve interpretation
REVE	2x	3.02e-21	Descriptive only
REVE	4x	6.57e-13	Descriptive only
LaBraM	2x	5.62e-20	Descriptive only
LaBraM	4x	1.62e-20	Descriptive only
CBraMod	2x	5.69e-12	Operational criterion met
CBraMod	4x	6.57e-13	Operational criterion met

Table S23: Completed population-budget decision records with the original six-slot positive-gain correction.

Model	Dataset	Budget	Delta raw p	U raw p	U budget Holm	Delta decline	U decline
REVE	Cho	1x	1.27e-08	2.57e-10	–	0.000	0.000
REVE	Dreyer	1x	6.75e-14	4.39e-15	–	0.013	0.000
REVE	PhysioNet	1x	1.26e-05	1.54e-13	–	0.107	0.029
REVE	Pooled	1x	2.1e-24	4.6e-36	1.38e-35	0.051	0.013
REVE	Cho	2x	2.31e-09	4.32e-09	–	0.000	0.000
REVE	Dreyer	2x	2.31e-13	3.92e-15	–	0.013	0.000

Table S24: All completed population-budget comparisons. Difference columns are in percentage points; decline columns are subject fractions.

Model	Dataset	Budget	CE mean (SD)	Accuracy % (SD)	BA % (SD)
CBraMod	Cho	1x	0.714 (0.145)	64.117 (6.690)	64.117 (6.690)
CBraMod	Dreyer	1x	0.531 (0.162)	81.178 (2.943)	81.178 (2.943)
CBraMod	PhysioNet	1x	0.839 (0.213)	81.957 (3.430)	82.013 (3.449)
CBraMod	Pooled	1x	0.709 (0.129)	77.733 (2.681)	77.758 (2.693)
CBraMod	Cho	2x	0.827 (0.224)	65.407 (5.218)	65.407 (5.218)
CBraMod	Dreyer	2x	0.643 (0.123)	82.239 (2.116)	82.239 (2.116)

Table S25: Validation performance at each completed population-budget endpoint. Run-level means and standard deviations use equal fold/seed weights.

Budget	Dataset	N	Seeds	Planned pairs	Observed pairs	Complete N	Incomplete N
1x	PhysioNet	103	5	515	494	82	21
1x	Dreyer	80	5	400	384	64	16
1x	Cho	52	5	260	249	41	11
1x	Pooled	235	5	1175	1127	187	48
2x	PhysioNet	103	5	515	494	82	21
2x	Dreyer	80	5	400	384	64	16

Table S26: Subject and seed coverage in the initial CBraMod budget batch before the documented rerun.

Seed	Budget	Status	N	G mean	Delta median	U median
11	1x	complete	48	75.774	4.083	5.208
11	2x	complete	48	77.521	4.083	5.483
11	4x	complete	48	77.868	2.083	3.826
23	1x	complete	48	76.888	3.750	5.958
23	2x	complete	48	78.172	0.833	4.178
23	4x	complete	48	77.488	3.894	5.140

Table S27: Initial CBraMod budget results by fold and seed. Missing personal diagnostics remain unavailable.

Fold	Seed	Budget	Dataset	CE	Accuracy (%)	BA (%)
1	37	1x	Cho	0.5447	66.50	66.50
1	37	1x	Dreyer	0.5782	79.44	79.44
1	37	1x	PhysioNet	1.0935	79.35	79.11
1	37	1x	Pooled	0.7998	76.52	76.42
1	37	2x	Cho	1.1068	67.00	67.00
1	37	2x	Dreyer	0.5750	82.64	82.64

Table S28: Available validation observations from the failed initial CBraMod run. Later endpoints are not filled.

Model	Budget	Initial six-slot Holm p	Initial interpretation
REVE	2x	3.02e-21	Descriptive only
REVE	4x	9.6e-13	Descriptive only
LaBraM	2x	5.62e-20	Descriptive only
LaBraM	4x	1.62e-20	Descriptive only
CBraMod	2x	NA	Incomplete cohort
CBraMod	4x	NA	Incomplete cohort

Table S29: Initial incomplete-batch budget decisions and their original six-slot correction, separate from the completed analysis.

Fold	Seed	All recorded events	Five labels	Ten labels	Twenty labels
0	11	0	0	0	0
0	23	0	0	0	0
0	37	0	0	0	0
0	53	0	0	0	0
0	71	0	0	0	0
1	11	0	0	0	0

Table S30: Recorded inner-loop numerical events by fold, seed and support-label count. Counts are events, not affected subjects.

Explore similar work

Universal BCI Personalization: One API for Frozen EEG Trunks and Foundation Models
Jul 24, 2026Sergey MusienkoElectroencephalography Foundation Models Brain-Computer Interface
Matched-Input Estimates Differ in Sign Across Architectures: Auditing EEG Foundation Models on Motor Imagery
Sep 20, 2026Kevin Zhou, Sparsh RoyElectroencephalography Foundation Models Brain-Computer Interface
Hopewell Valley Central High School
Channel Adaptation for EEG Foundation Models: A Systematic Benchmark Across Architectures, Tasks, and Training Regimes
Apr 25, 2026Kuntal Kokate, Bruno Aristimunha, Dung Truong +1Electroencephalography Foundation Models Parameter-Efficient Adaptation
Swartz Center for Computational Neuroscience, Institute for Neural Computation, University of California San Diego, La Jolla, CA 92093, USA · Yneuro, Paris, France · Centre de Recherche Cerveau et Cognition, CNRS, Université Paul Sabatier, Toulouse, France

Jul 24, 2026, cs.HCJ/K move · Enter open · S save

Universal BCI Personalization: One API for Frozen EEG Trunks and Foundation Models

Sergey Musienko

Electroencephalography Foundation Models Brain-Computer Interface Electroencephalography Model Fine-Tuning Personalization Encoders+3

Open paper AskPDF arXiv

Frozen EEG encoders proliferate; per-model fine-tune defaults do not scale. We present Nimbus Personalizer: one contract encode to Bayesian head to BrainState (optional affine mid-tier) that sits on heterogeneous frozen trunks without a new personalization stack per architecture. Thesis (systems): the contribution is the trunk-agnostic API - not LDA-on-embeddings as an ML novelty - so OEMs integrate once and swap trunks. Evidence: the same surface runs on five classical trunks EEGNet, Shallow, Deep, Conformer, ATCNet x four MI datasets (18 cells) and on a foundation encoder (REVE) under the same Personalizer. Where embedding capacity exists, the head is a cheap default mid-point versus warm-start fine-tune or PEFT, costing orders of magnitude less adaptation wall time while recovering much of the fine-tune accuracy gain; calibration-only-when-clean holds in 12/18 cells. Head gains are supporting evidence that the API is useful where capacity exists. Subject-level confidence intervals identify the clearest dataset and span zero elsewhere. All results are exploratory (subject-level bootstrap, no confirmatory tests); the decision logic for when to escalate adaptation is addressed in our companion work on the control layer.

Abstract

Explore similar work

Universal BCI Personalization: One API for Frozen EEG Trunks and Foundation Models

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

Channel Adaptation for EEG Foundation Models: A Systematic Benchmark Across Architectures, Tasks, and Training Regimes