CoChat

Transfer learning for EEG-based BCIs: a comparative evaluation and optimization of data alignment methods

Open access
Ahmed, Soha Galalaldin, Ambali Parambil, Medha Mohan, Damseh, Rafat, Bouktif, Salah, Alnajjar, Fady, Belkacem, Abdelkader Nasreddine · 2026 · Frontiers in Systems Neuroscience

The DOI matches its Crossref record; 51 of the 51 references it cites were found in a registry or an index; 9 of the 9 sources its claims cite were checked for support.

Abstract

Background: This paper addresses a critical challenge in developing practical EEG-based brain-computer interfaces (BCIs): enhancing cross-subject generalization by mitigating individual differences in brain signals. How can we effectively leverage data from existing subjects to improve performance for a new user with minimal subject-specific calibration? Methods: We systematically compare and optimize three prominent data alignment techniques, Riemannian Procrustes Analysis (RPA), Euclidean Alignment (EA), and Correlation Alignment (CORAL), designed to transform EEG data from multiple source subjects and a target subject into a common representation space, mitigating variability. Evaluation: We employed leave-one-subject-out cross-validation (LOSO-CV) framework on EEG-based attention decoding data to empirically evaluate the effectiveness of each alignment method compared to a baseline condition with no alignment. Key parameters, specifically the regularization parameter α for EA, were optimized to maximize cross-subject transfer performance. Results: ≈0.59; 95% confidence interval for the mean improvement [0.52%, 6.36%]). Because this α value was identified from the same sweep that produced the per-subject accuracies, this estimate together with the per-subject "best-parameter" results should be interpreted as an oracle sensitivity-analysis upper bound on subject-specific tuning rather than as a leakage-free LOSO estimate. While optimized EA showed the best mean performance, the analysis also demonstrated subject-specific differences in the most ideal alignment strategy. Conclusion: This comparison framework quantifies the benefits of different alignment approaches and highlights the valuable contribution of parameter optimization, particularly for EA. Significance: These results indicate the potential of optimized alignment techniques, EA in particular, to significantly enhance cross-subject transfer learning in EEG-based BCIs. This has practical ramifications for methodology selection and tuning, and maps a path toward more robust and generalizable BCI systems requiring less subject-specific calibration for real-world applications.

Overview

CoChat Summary

TL;DR

This paper comparatively evaluates three EEG data-alignment methods—RPA, EA, and CORAL—for cross-subject transfer in an attention-decoding BCI using leave-one-subject-out cross-validation. Optimized EA at α = 100 gave the largest mean improvement over no alignment, but the authors emphasize that per-subject tuning results are oracle upper bounds rather than leakage-free generalization estimates.

Key findings

  • Alignment improved classification over the no-alignment baseline; EA at α = 100 yielded a mean improvement of 3.44% (SD = 5.88), paired t(17) ≈ 2.48, p ≈ 0.024, Cohen’s dz ≈ 0.59, 95% CI [0.52%, 6.36%].
  • CORAL was frequently the best alignment method, appearing as the top method for 9 out of 18 subjects, while EA often appeared as second-best.
  • Parameter sensitivity was substantial: CORAL generally performed best at low λ (≤ 0.1), while EA often favored high α (e.g., α = 100) and declined sharply beyond α = 1000.
  • RPA’s best performance occurred when α was close to 1, which the authors interpret as evidence that RPA may not provide effective alignment because covariance structure is largely replaced by an identity matrix.

Methods

  • Used the Reichert et al. (2020) covert spatial attention EEG dataset: 18 healthy participants, 14 parieto-occipital channels, 168 trials per subject, with preprocessing including mastoid re-referencing, 1.0–12.5 Hz band-pass filtering, 50 Hz resampling, and 750 ms epochs.
  • Compared No Alignment, RPA, EA, and CORAL within a LOSO-CV framework; decoding used Canonical Correlation Analysis (CCA), and alignment parameters λ for CORAL and α for EA were swept over [10⁻⁴, 10⁴].
  • Evaluated per-subject and mean decoding accuracy, with paired t-tests against the no-alignment baseline and subject-specific best-parameter analyses.

Limitations

  • Stated: evaluation is restricted to a single covert spatial attention dataset with 18 subjects, limiting generalizability to other paradigms such as motor imagery or alternative ERP tasks.
  • Stated: the study does not benchmark recent deep transfer, self-supervised, or foundation-model approaches.
  • Stated: per-subject parameter sweeps and the fixed α = 100 result are not leakage-safe because parameters were selected using the held-out target subject’s accuracy; they are oracle upper bounds, and a leakage-free evaluation would require nested LOSO with parameter selection restricted to source subjects.
  • Stated: the 1.0–12.5 Hz pass-band excludes higher-frequency activity that may be informative in other paradigms.
  • Evident from text: no wall-clock timings are reported, as the authors state the goal was asymptotic behavior rather than hardware-specific benchmarks.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3389/fnsys.2026.1840121

Cites

51 of 51 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Taxonomy of transfer learning in BCI: This framework categorizes intra-subject and inter-subject transfer approaches.
Figure 1 Taxonomy of transfer learning in BCI: This framework categorizes intra-subject and inter-subject transfer approaches.CC BY · PubMed Central
Figure 2 Methodological pipeline of the study: the stepwise process of the study, including EEG pre-processing, label preparation, alignment methods, spatial filtering, classification, and performance evaluation using LOSO validation.
Figure 2 Methodological pipeline of the study: the stepwise process of the study, including EEG pre-processing, label preparation, alignment methods, spatial filtering, classification, and performance evaluation using LOSO validation.CC BY · PubMed Central
Figure 3 Electrode placements based on the 10–20 system, highlighting the 14 parieto-occipital channels selected for attention decoding in this study.
Figure 3 Electrode placements based on the 10–20 system, highlighting the 14 parieto-occipital channels selected for attention decoding in this study.CC BY · PubMed Central
Figure 4 Experiment procedure.
Figure 4 Experiment procedure.CC BY · PubMed Central
Figure 5 Decoding accuracy per subject for each alignment method. The figure presents accuracy for Subject-Dependent and Transfer Learning approaches including No Alignment, RPA, EA, and CORAL. The “RPA”, “EA”, and “CORAL” bars use each method's default parameterisation and are leakage-free LOSO estimates. Per-subject tuned accuracies (Figure 7, Tables 4, 5) are oracle upper bounds.
Figure 5 Decoding accuracy per subject for each alignment method. The figure presents accuracy for Subject-Dependent and Transfer Learning approaches including No Alignment, RPA, EA, and CORAL. The “RPA”, “EA”, and “CORAL” bars use each method's default parameterisation and are leakage-free LOSO estimates. Per-subject tuned accuracies (Figure 7, Tables 4, 5) are oracle upper bounds.CC BY · PubMed Central
Figure 7 Per-subject oracle accuracies for EA and CORAL: for each subject and each method, the curve shows the highest accuracy attained across the full parameter sweep, with the corresponding λ or α annotated. Because the parameter is chosen by inspecting the target subject's own decoding accuracy, these values are an upper bound on what subject-specific tuning could achieve and are not a leakage-safe estimate of generalization to a new subject. Comparison of best accuracy achieved by EA Optimization (blue line with circular markers) and CORAL Optimization (red line with square markers) across 18 subjects. The x-axis shows subject IDs and the y-axis shows accuracy percentages. Optimal parameter values for each subject are annotated.
Figure 7 Per-subject oracle accuracies for EA and CORAL: for each subject and each method, the curve shows the highest accuracy attained across the full parameter sweep, with the corresponding λ or α annotated. Because the parameter is chosen by inspecting the target subject's own decoding accuracy, these values are an upper bound on what subject-specific tuning could achieve and are not a leakage-safe estimate of generalization to a new subject. Comparison of best accuracy achieved by EA Optimization (blue line with circular markers) and CORAL Optimization (red line with square markers) across 18 subjects. The x-axis shows subject IDs and the y-axis shows accuracy percentages. Optimal parameter values for each subject are annotated.CC BY · PubMed Central
Figure 6 Mean leave-one-subject-out classification accuracy across alignment settings. Bars show the mean across 18 subjects and error bars show standard deviation. The dashed line marks chance level at 50%. Asterisks indicate paired comparisons against No Alignment. Methods: (A) No Alignment, (B) RPA, (C) EA optimization (α = 100), (D) Optimized CORAL (per-subject best λ), (E) Optimized EA (per-subject best α), (F) EA, (G) CORAL, (H) CORAL optimization (λ = 0.001).
Figure 6 Mean leave-one-subject-out classification accuracy across alignment settings. Bars show the mean across 18 subjects and error bars show standard deviation. The dashed line marks chance level at 50%. Asterisks indicate paired comparisons against No Alignment. Methods: (A) No Alignment, (B) RPA, (C) EA optimization (α = 100), (D) Optimized CORAL (per-subject best λ), (E) Optimized EA (per-subject best α), (F) EA, (G) CORAL, (H) CORAL optimization (λ = 0.001).CC BY · PubMed Central

Verification

Claims the paper makes

7 Valid · 1 Not valid · 0 Not checked · 83 Own results and interpretations

50 more claims

Create a free account to see all 91 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 29
  • Probably not a substantial claim. 11

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

3 of 3

Citations

Cites

Research works printed but not linked: 2

  1. Alamgir, 2010, “Multitask learning for brain-computer interfaces,”, 17
  2. Honzík V. Ph.D. thesis Use of spiking neural networks 2021

Metadata

Authors
Ahmed, S. G., Ambali Parambil, M. M., Damseh, R., Bouktif, S., Alnajjar, F., & Belkacem, A. N.
Affiliations
  • United Arab Emirates University (AE)
  • Department of Computer Science and Software Engineering, College of Information Technology, United Arab Emirates University
Corresponding author
Alnajjar, Fady
Venue
Frontiers in Systems Neuroscience
Published
2026-06-11
Volume
20
Pages
1840121-1840121
ISSN
1662-5137 (ISSN-L)
OpenAlex
W7164296090
Cited by
2
References
51
Citations by year
2026: 2
FWCI
8.24
Type
journal article
Language
en
Open access
Locations
  • Frontiers in Systems Neuroscience (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
United Arab Emirates University
Topics
  • Life Sciences › Neuroscience › Cognitive Neuroscience › EEG and Brain-Computer Interfaces
  • Physical Sciences › Computer Science › Human-Computer Interaction › Gaze Tracking and Assistive Technology
  • Physical Sciences › Computer Science › Artificial Intelligence › Emotion and Mood Recognition
Keywords
EEG-based brain-computer interface, cross-subject transfer learning, data alignment, coral, leave-one-subject-out cross-validation, regularization parameter optimization, subject-specific calibration, interindividual variability

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.