# BCI-002: A review of classification algorithms for EEG-based brain-computer interfaces: a 10 year update > Generated from a local paper card. Do not edit this page directly; edit the source card and rerun `vp run docs:generate`. > Internal wiki boundary: paper-card evidence and user-provided PDF downloads are available for private reading. Extracted full-paper text is not published. > Source: `library/paper_cards/BCI-002.md` ## Paper Access * Internal PDF: Download Paper * DOI / official page: [10.1088/1741-2552/aab2f2](https://doi.org/10.1088/1741-2552/aab2f2) * Deployment boundary: these PDF links are intended only for a private/protected internal wiki. ## BCI-002: A review of classification algorithms for EEG-based brain-computer interfaces: a 10 year update ## Metadata * ID: BCI-002 * Title: A review of classification algorithms for EEG-based brain-computer interfaces: a 10 year update * Year: 2018 * DOI / URL: 10.1088/1741-2552/aab2f2 * Local PDF: see Paper Access section above * Text artifact: local-only path withheld from docs site * Review status: `extracted` ## Study Type * Track: BCI / EEG Foundations * Task: topical review of EEG-based BCI classification algorithms published mainly from 2007 to 2017 * Participants or dataset: review paper; no single participant cohort * Hardware: EEG-based BCI systems across reviewed studies * Channels or sensors: EEG, with channel layouts depending on the reviewed BCI paradigm ## Methods * Paradigm: survey and synthesis of machine-learning methods for EEG BCI classification * Signal processing or model: adaptive classifiers, matrix/tensor classifiers, transfer learning, deep learning, shrinkage LDA, random forests, Riemannian geometry classifiers * Training/calibration: emphasizes offline training, online adaptation, transfer learning, and calibration reduction * Online/offline: review distinguishes offline evaluations from online BCI operation and argues online validation is necessary ## Results * Metrics: classification performance, online usability, computational efficiency, calibration burden, robustness to noise and limited training data * Main findings: adaptive classifiers generally outperform static classifiers; Riemannian geometry methods are state-of-the-art across multiple BCI problems; shrinkage LDA, random forest, and Riemannian classifiers are useful with limited training data; transfer learning can help but is not reliably beneficial; deep learning had not yet shown consistent improvement over state-of-the-art BCI methods in this review * Reported limitations: many classifiers are evaluated only offline; online robustness, calibration time, real-life noise, and user feedback remain open challenges ## Relevance To This Project * Supports: decoder baseline selection, online-validation caution, calibration-burden framing, and the need to log offline versus online claims separately * Conflicts with: does not evaluate SAH-BRI-Grasp, dynamic object SSVEP targets, MI stop/cancel commands, or robot manipulation * Design implication: use sLDA/Riemannian/CSP-family and compact deep models as baselines where appropriate, but require online Exp1/Exp2 validation before claiming operational BCI reliability ## Extracted Evidence | Claim | Status | Evidence Note | Page/Section | | --- | --- | --- | --- | | EEG BCI classification methods must be evaluated online, not only offline, to support real BCI claims. | verified | The review states that actual BCI applications are fundamentally online and that online classifier evaluation should be the norm. | Section 5.2 | | Adaptive classifiers should generally be preferred over static classifiers when feasible. | verified | The guidelines state that adaptive classifiers and adaptive spatial filters generally outperform static ones, including unsupervised adaptation. | Section 5.1; Conclusion | | Riemannian geometry classifiers are a strong current baseline for multiple EEG BCI paradigms. | verified | The review identifies Riemannian classifiers as current state-of-the-art for several BCI problems, including MI, P300, and SSVEP classification. | Section 5.1 | | Deep learning should not be assumed superior for EEG BCI without sufficient data and comparison. | verified | The review says deep learning methods were lagging in performance for BCI, largely due to limited training datasets, while shallow CNNs may be more promising. | Abstract; Section 5.1 | | SAH-BRI-Grasp should separate offline decoder evidence from online control evidence. | inferred | This follows from the paper's online-validation warnings, but the review does not evaluate SAH-BRI-Grasp. | Sections 5.1-5.2 | ## Open Questions * Which decoder baselines should be fixed for Exp1 and Exp2 after considering remaining MI/SSVEP cards? * Should Exp2 compare CSP/FBCSP, Riemannian, sLDA, and EEGNet under the same no-control protocol? * How much online calibration time is acceptable for the target user workflow?