# MI-016: Multiclass Brain-Computer Interface Classification by Riemannian Geometry

> 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-016.md`

## Paper Access

* Internal PDF: <a href={"/papers/MI-016.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: [10.1109/TBME.2011.2172210](https://doi.org/10.1109/TBME.2011.2172210)
* Deployment boundary: these PDF links are intended only for a private/protected internal wiki.

## MI-016: Multiclass Brain-Computer Interface Classification by Riemannian Geometry

## Metadata

* ID: MI-016
* Title: Multiclass Brain-Computer Interface Classification by Riemannian Geometry
* Year: 2012
* DOI / URL: 10.1109/TBME.2011.2172210
* Local PDF: see Paper Access section above
* Text artifact: local-only path withheld from docs site
* Review status: `extracted`

## Study Type

* Track: MI
* Task: multiclass BCI classification with Riemannian geometry on covariance matrices
* Participants or dataset: BCI Competition IV Dataset IIa
* Device/electrode setup: dataset EEG; exact channel montage follows BCI Competition IV IIa
* Protocol/task: multiclass motor-imagery classification

## Methods

* Signal processing or analysis: covariance matrices on the SPD manifold; minimum distance to mean and tangent-space LDA
* Comparator: multiclass CSP + LDA
* Online/offline: offline benchmark evaluation

## Key Results

* The abstract reports improvement in mean classification accuracy from 65.1% to 70.2%.
* Riemannian methods reduce the need to separate spatial filtering and classification into distinct steps.

## Limitations

* Offline dataset evaluation only.
* Does not directly address asynchronous no-control or robot interventions.
* Dataset-specific results should not be overgeneralized.

## Relevance To Current Review

* Introduces the Riemannian branch of MI decoding.
* Important for recent transfer learning and geometry-based EEG trends.

## Evidence Status

| Claim | Status | Evidence Note |
| --- | --- | --- |
| Riemannian geometry treats EEG covariance matrices as manifold-valued features. | verified | Abstract and methods describe covariance matrices and the Riemannian manifold. |
| The paper improved mean accuracy over a CSP/LDA comparator on Dataset IIa. | verified | Abstract reports 65.1% to 70.2%. |
| Riemannian methods guarantee online no-control safety. | needs confirmation | Online safety was not evaluated. |

## Open Questions

* Should Exp2 compare CSP/FBCSP against a minimum-distance-to-mean Riemannian baseline?
* How many channels are required for stable covariance estimation in low-channel settings?