# MI-007: Filter Bank Common Spatial Pattern (FBCSP) in Brain-Computer Interface

> 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/MI-007.md`

## Paper Access

* Internal PDF: <a href={"/papers/MI-007.pdf"} download style={{ display: "inline-flex", alignItems: "center", justifyContent: "center", minHeight: "2.25rem", padding: "0.45rem 0.8rem", borderRadius: "6px", backgroundColor: "#047857", color: "#ffffff", fontWeight: 700, lineHeight: 1, textDecoration: "none", boxShadow: "0 1px 2px rgba(15, 23, 42, 0.22)" }}>Download Paper</a>
* DOI / official page: [10.3389/fnins.2012.00039](https://doi.org/10.3389/fnins.2012.00039)
* Open-access page: [Open access page](https://www.frontiersin.org/articles/10.3389/fnins.2012.00039/full)
* Deployment boundary: these PDF links are intended only for a private/protected internal wiki.

## MI-007: Filter Bank Common Spatial Pattern (FBCSP) in Brain-Computer Interface

## Metadata

* ID: MI-007
* Title: Filter Bank Common Spatial Pattern (FBCSP) in Brain-Computer Interface
* Year: 2012
* DOI / URL: 10.3389/fnins.2012.00039
* Local PDF: see Paper Access section above
* Text artifact: local-only path withheld from docs site
* Review status: `extracted`

## Study Type

* Track: MI
* Task: evaluate FBCSP on BCI Competition IV Datasets 2a and 2b
* Participants or dataset: Dataset 2a has 4-class motor imagery EEG from 9 subjects; Dataset 2b has 2-class EEG from 9 subjects
* Hardware: competition datasets; Dataset 2a uses 22 EEG channels, Dataset 2b uses 3 bipolar EEG channels over C3/Cz/C4 plus EOG
* Channels or sensors: EEG motor imagery recordings with multiple sessions and evaluation data

## Methods

* Paradigm: motor imagery classification for left hand, right hand, feet, and tongue in Dataset 2a; two-class MI in Dataset 2b
* Signal processing or model: FBCSP with Chebyshev Type II filter bank, CSP spatial filters, mutual-information feature selection, and Bayesian classification; multi-class extensions include divide-and-conquer, pair-wise, and one-versus-rest
* Training/calibration: 10 x 10-fold cross-validation on training data and session-to-session transfer on evaluation data
* Online/offline: offline benchmark evaluation on competition datasets

## Results

* Metrics: single-trial classification accuracy and Cohen's kappa
* Main findings: FBCSP performed relatively best among competition submissions after labels were disclosed, with mean kappa 0.569 on Dataset 2a and 0.600 on Dataset 2b; OVR and PW were strong multi-class extensions for Dataset 2a
* Reported limitations: benchmark offline evaluation; FBCSP still depends on training data, channel quality, and feature-selection choices; not tested for SAH-BRI-Grasp no-control or robot tasks

## Relevance To This Project

* Supports: FBCSP as a strong MI baseline and low-channel benchmark reference for Exp2/Exp4
* Conflicts with: does not evaluate asynchronous no-control states, stop/cancel commands, or shared-autonomy robot control
* Design implication: Exp2 should include FBCSP as a baseline for MI mode commands, especially when channel subsets are tested

## Extracted Evidence

| Claim | Status | Evidence Note | Page/Section |
| --- | --- | --- | --- |
| FBCSP addresses subject-specific MI frequency-band selection. | verified | The introduction states that FBCSP optimizes the subject-specific frequency band for CSP. | Abstract; Introduction |
| FBCSP can handle BCI Competition IV 4-class and 2-class MI datasets. | verified | The methods describe Dataset 2a 4-class MI and Dataset 2b 2-class MI with corresponding FBCSP extensions. | Methods |
| FBCSP achieved strong competition performance on both datasets. | verified | The abstract reports mean kappa values of 0.569 and 0.600 on Datasets 2a and 2b and says it performed relatively best among submissions. | Abstract; Results |
| Dataset 2b provides a low-channel MI benchmark using 3 bipolar channels. | verified | The dataset description states Dataset 2b used 3 bipolar recordings at C3, Cz, and C4. | Methods |
| SAH-BRI-Grasp can use FBCSP as an MI decoder baseline, but project safety claims require no-control tests. | inferred | The benchmark does not evaluate no-control false activation or robot execution. | Methods; Results |

## Open Questions

* Which FBCSP feature-selection method should be used for the first Exp2 baseline?
* How will FBCSP compare against EEGNet and simpler CSP under the project channel montage?
* Can FBCSP be made stable enough for online pause/stop/cancel use?