Direct information transfer rate optimisation for SSVEP-based BCI

Anti Ingel; Ilya Kuzovkin; Raul Vicente

arxiv: 1907.10509 · v1 · pith:KVIMOJ2Unew · submitted 2019-07-19 · 📡 eess.SP · cs.LG· stat.ML

Direct information transfer rate optimisation for SSVEP-based BCI

Anti Ingel , Ilya Kuzovkin , Raul Vicente This is my paper

Pith reviewed 2026-05-24 18:44 UTC · model grok-4.3

classification 📡 eess.SP cs.LGstat.ML

keywords SSVEPBCIinformation transfer ratethreshold optimizationclassificationbrain-computer interfacesteady-state visually evoked potential

0 comments

The pith

Optimizing per-feature thresholds by directly maximizing a general ITR formula doubles SSVEP BCI performance to 62 bits per minute.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces a classification method for steady-state visually evoked potential brain-computer interfaces that starts with standard feature extraction but then automatically tunes discrimination thresholds for each feature. The tuning is performed by maximizing information transfer rate, for which the authors derive a general formula that remains accurate even when the usual assumptions about the data no longer hold. Because the optimization is direct, the method removes any need for manual threshold selection or exhaustive grid searches over parameters. When tested on the dataset the approach reaches an ITR of 62 bits per minute, twice the value reported by earlier methods, and simultaneously lowers the rate of false classifications.

Core claim

By replacing the standard ITR formula with a more general expression that does not rely on its usual assumptions, the authors can treat the selection of discrimination thresholds as a direct maximisation problem; solving that problem for each feature yields thresholds that maximise measured performance without manual tuning or grid search and produce an ITR of 62 bit/min on the evaluated dataset.

What carries the argument

The general ITR formula, used as the objective function to optimise discrimination thresholds for each extracted SSVEP feature.

If this is right

Thresholds for any set of SSVEP features can be found automatically by maximising the general ITR expression.
The occurrence of false classifications is reduced without additional post-processing steps.
No manual parameter selection or computationally expensive grid search is required to reach the reported performance level.
On the dataset examined the resulting ITR is 62 bit/min, twice the value of previously published results.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same optimisation procedure could be applied to feature sets from other BCI signal types where manual threshold tuning is currently common.
Lower false-positive rates may improve safety margins when the interface is used for real-time control tasks.
The general ITR formula offers a more reliable way to compare classification pipelines across studies that differ in trial length or error statistics.
Retraining the thresholds on each new user session would test how much the reported gains depend on session-specific data.

Load-bearing premise

The general ITR formula correctly measures true information transfer when standard assumptions fail, and thresholds tuned on the available data continue to work for new users or sessions.

What would settle it

Apply the automatically chosen thresholds to a fresh recording session from a new user and check whether the realised ITR matches the value predicted by the general formula or falls substantially below it.

read the original abstract

In this work, a classification method for SSVEP-based BCI is proposed. The classification method uses features extracted by traditional SSVEP-based BCI methods and finds optimal discrimination thresholds for each feature to classify the targets. Optimising the thresholds is formalised as a maximisation task of a performance measure of BCIs called information transfer rate (ITR). However, instead of the standard method of calculating ITR, which makes certain assumptions about the data, a more general formula is derived to avoid incorrect ITR calculation when the standard assumptions are not met. This allows the optimal discrimination thresholds to be automatically calculated and thus eliminates the need for manual parameter selection or performing computationally expensive grid searches. The proposed method shows good performance in classifying targets of a BCI, outperforming previously reported results on the same dataset by a factor of 2 in terms of ITR. The highest achieved ITR on the used dataset was 62 bit/min. The proposed method also provides a way to reduce false classifications, which is important in real-world applications.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper derives a general ITR formula to auto-optimize per-feature thresholds in SSVEP BCIs and reports doubling ITR to 62 bit/min, but the optimization on the full dataset without clear cross-validation leaves the gains vulnerable to optimistic bias.

read the letter

The main thing here is a procedure that turns threshold selection into a direct maximization of a generalized ITR formula instead of grid search or hand tuning. They start from standard SSVEP features, derive a broader ITR expression that drops some of the usual assumptions, and use it to pick discrimination thresholds per feature. On the dataset they test, this yields 62 bit/min, roughly twice the prior reported numbers, and they note it also helps cut false positives. That part is straightforward engineering and removes a common pain point in BCI pipelines. The derivation itself is presented as new, and the automatic aspect is a clear practical win if it works out of sample. The paper is aimed at people who already run SSVEP systems and want fewer free parameters. A reader who builds or tunes these classifiers could pick up the threshold method and test it themselves. The soft spot is exactly the one the stress-test flags. Thresholds are chosen to maximize the same ITR quantity later reported as the result, and nothing in the abstract or the described experiments indicates held-out validation or k-fold checks. That setup can inflate performance on the training statistics alone. Without seeing the full derivation and the experimental splits, it is hard to tell how much of the 2x gain survives on new users or sessions. The circularity is not fatal on its own, but it is the load-bearing claim, so any referee would need to see explicit checks on that point. The math and data handling look like honest work within the subfield, but the evidence for generalization is thin so far. I would send it to peer review so the authors can supply the missing validation steps and the actual formula; the topic is narrow enough that a focused referee could settle whether the gains are real or in-sample artifacts.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes a classification method for SSVEP-based BCIs that uses traditionally extracted features and optimizes per-feature discrimination thresholds by directly maximizing a derived general information transfer rate (ITR) formula. This replaces the standard ITR calculation (which relies on assumptions that may not hold) and is claimed to enable automatic threshold selection without manual tuning or grid search. On the evaluated dataset the method reportedly achieves a peak ITR of 62 bit/min, doubling previously published results on the same data, while also reducing false classifications.

Significance. A correctly derived general ITR expression that remains valid when standard assumptions fail would be a useful contribution for BCI performance assessment. If the optimization procedure generalizes beyond the training samples, the reported doubling of ITR and the elimination of manual parameter search could have practical value for real-world SSVEP systems. The manuscript does not, however, supply machine-checked derivations, reproducible code, or parameter-free results that would strengthen the claim.

major comments (3)

[Methods (threshold optimization procedure) and Results (ITR reporting)] The central performance claim (62 bit/min ITR and 2× improvement) rests on thresholds that are selected by maximizing the same ITR quantity later reported as the result. No cross-validation, held-out test set, or per-subject/session partitioning is described for the optimization step; this makes the reported gain vulnerable to in-sample overfitting of dataset-specific statistics.
[Derivation of the general ITR formula (presumably §3 or §4)] The abstract states that a general ITR formula was derived, yet supplies neither the expression itself nor any validation against the standard ITR formula on synthetic or real data where the usual assumptions hold. Without these steps it is impossible to confirm that the new formula supplies independent grounding rather than re-expressing the fitted performance.
[Results (performance comparison)] Table or figure reporting the 62 bit/min result and the 2× comparison does not indicate whether the comparison baselines were also evaluated under identical train/test splits or whether they used the same feature extraction pipeline; this undermines the direct factor-of-two claim.

minor comments (2)

[Abstract] The abstract contains no equations or algorithmic outline; a short pseudocode box or one-line statement of the optimization objective would improve readability.
[Notation and Methods] Notation for the per-feature thresholds and the general ITR expression should be introduced consistently in the methods section and reused in the results.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments and the opportunity to clarify our work. We respond to each major comment below, indicating where revisions will be made.

read point-by-point responses

Referee: [Methods (threshold optimization procedure) and Results (ITR reporting)] The central performance claim (62 bit/min ITR and 2× improvement) rests on thresholds that are selected by maximizing the same ITR quantity later reported as the result. No cross-validation, held-out test set, or per-subject/session partitioning is described for the optimization step; this makes the reported gain vulnerable to in-sample overfitting of dataset-specific statistics.

Authors: We agree this is a valid concern. The optimization of thresholds via the general ITR is performed on the dataset to illustrate the method, but the lack of explicit cross-validation or partitioning leaves the results open to overfitting criticism. We will revise the Methods section to incorporate a cross-validation scheme (e.g., leave-one-session-out) for threshold selection and report the resulting ITR values separately from the in-sample optimization. revision: yes
Referee: [Derivation of the general ITR formula (presumably §3 or §4)] The abstract states that a general ITR formula was derived, yet supplies neither the expression itself nor any validation against the standard ITR formula on synthetic or real data where the usual assumptions hold. Without these steps it is impossible to confirm that the new formula supplies independent grounding rather than re-expressing the fitted performance.

Authors: The general ITR formula and its derivation appear in Section 3. We acknowledge that the abstract omits the explicit expression and that no direct validation is provided. We will revise the abstract to include the formula and add a validation subsection comparing the general ITR to the standard formula on synthetic data (where assumptions hold) as well as the real dataset. revision: yes
Referee: [Results (performance comparison)] Table or figure reporting the 62 bit/min result and the 2× comparison does not indicate whether the comparison baselines were also evaluated under identical train/test splits or whether they used the same feature extraction pipeline; this undermines the direct factor-of-two claim.

Authors: The factor-of-two claim is based on published results for the same dataset. To strengthen transparency, we will revise the Results section to explicitly state the evaluation conditions used for baselines, including feature extraction and data usage. Where possible, we will note any differences in protocol. revision: partial

Circularity Check

1 steps flagged

Thresholds optimized by maximizing the reported ITR quantity on the evaluation data

specific steps

fitted input called prediction [Abstract]
"Optimising the thresholds is formalised as a maximisation task of a performance measure of BCIs called information transfer rate (ITR). ... The proposed method shows good performance in classifying targets of a BCI, outperforming previously reported results on the same dataset by a factor of 2 in terms of ITR. The highest achieved ITR on the used dataset was 62 bit/min."

Thresholds are selected by maximizing ITR on the dataset; the reported ITR (62 bit/min) is then the value attained by those same thresholds on the identical data. The performance number is therefore the direct output of the fitting procedure rather than a separate evaluation.

full rationale

The paper formalizes threshold selection as direct maximization of the (generalized) ITR on the same samples later used to report final ITR performance. This matches the fitted-input-called-prediction pattern: the quantity presented as the result is the objective that was maximized during parameter selection. No cross-validation or held-out set is indicated in the provided text, so the 62 bit/min figure and 2× improvement claim are the in-sample optimum rather than an independent test of the formula. The derivation of the general ITR formula itself is not shown to be circular, but the load-bearing performance claim reduces to the optimization step.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Review performed on abstract only; no explicit free parameters, axioms, or invented entities are stated in the provided text.

pith-pipeline@v0.9.0 · 5713 in / 990 out tokens · 38927 ms · 2026-05-24T18:44:14.460806+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

26 extracted references · 26 canonical work pages

[1]

Brain-compu ter interface is a direct communication channel between the brain and some external devic e

Introduction In this work, a classiﬁcation method for steady-state visual evok ed potential (SSVEP) based brain-computer interface (BCI) is proposed. Brain-compu ter interface is a direct communication channel between the brain and some external devic e. SSVEP-based BCI uses visual stimuli, called targets, to elicit a response in the brain of t he user an...

work page
[2]

SSVEP dataset for oﬄine experiments The performance of the proposed method was tested using publicly available SSVEP dataset by Bakardijan et al

Methods 2.1. SSVEP dataset for oﬄine experiments The performance of the proposed method was tested using publicly available SSVEP dataset by Bakardijan et al. ‡ which contains SSVEP data from four subjects [8]. The dataset contains EEG recordings of the subjects’ brain respons es to 8 Hz, 14 Hz and 28 Hz ﬂickering. The visual stimuli were displayed on a c...

work page
[3]

Classiﬁcation method This section describes the derivation of the proposed classiﬁcation method that is based on ITR maximisation

Results 3.1. Classiﬁcation method This section describes the derivation of the proposed classiﬁcation method that is based on ITR maximisation. It follows the description of the method g iven in [14]. First, the classiﬁcation rule is described, and then a formula for est imating ITR from discrimination thresholds is derived. Finally, it is shown that the ...

work page
[4]

BIFB No ML 3 N/A 8.4 88% 8.2 -

work page
[5]

PSDA Bilinear separation 2 1 N/A 74% 16 -

work page
[6]

This can be seen as a trade-oﬀ between accuracy and mean detection time: allowing the c lassiﬁer to not make predictions results in better accuracy but worse mean detection t ime

Discussion From the results it can be seen that allowing the classiﬁer to not make predictions when it is not conﬁdent enough results in better performance. This can be seen as a trade-oﬀ between accuracy and mean detection time: allowing the c lassiﬁer to not make predictions results in better accuracy but worse mean detection t ime. According to ITR, wh...

work page
[7]

Direct information transfer rate optimisation for SSVEP-b ased BCI 10 • Combining feature extraction methods outperforms the same met hods used individually (Table 1)

Conclusion The main conclusions of the work are the following • Proposed classiﬁer outperforms previous results (Table 3). Direct information transfer rate optimisation for SSVEP-b ased BCI 10 • Combining feature extraction methods outperforms the same met hods used individually (Table 1). • Allowing the classiﬁer to not make predictions when it is not co...

work page
[8]

We also acknowledge fund ing by the European Regional Development Fund through the Estonian Center of Excelle nce in IT, EXCITE

Acknowledgments The authors thank the ﬁnancial support from The Estonian Resea rch Council through the personal research grant PUT1476. We also acknowledge fund ing by the European Regional Development Fund through the Estonian Center of Excelle nce in IT, EXCITE

work page
[9]

Zhu D, Bieger J, Molina G G and Aarts R M 2010 A survey of stimulation methods used in SSVEP-based BCIs Computational Intelligence and Neuroscience 2010 1–13

work page 2010
[10]

Wolpaw J R, Ramoser H, McFarland D J and Pfurtscheller G 1998 EEG -Based Communication: Improved Accuracy by Response Veriﬁcation IEEE Transactions on Rehabilitation Engineering 6 326–333

work page 1998
[11]

Jukiewicz M and Cysewska-Sobusiak A 2015 Implementation of Biline ar Separation algorithm as a classiﬁcation method for SSVEP-based brain-computer interfac e Measurement Automation Monitoring 61 51–53

work page 2015
[12]

Karnati V B R, Verma U S G, Amerineni J S and Shah V H 2014 Freque ncy detection in Medium and High frequency SSVEP based Brain Computer Interface syste ms by scaling of sine-curve ﬁt amplitudes 2014 First International Conference on Networks Soft Compu ting (ICNSC2014) pp 300–303

work page 2014
[13]

Velchev Y, Radev D and Radeva S 2016 Features Extraction Base d on Subspace Methods with Application to SSVEP BCI International Journal of Emerging Engineering Research an d Technology 4 52–58

work page 2016
[14]

Anindya S F, Rachmat H H and Sutjiredjeki E 2016 A prototype of SSVEP-based BCI for home appliances control 2016 1st International Conference on Biomedical Engineeri ng (IBIOMED) pp 1–6

work page 2016
[15]

Demir A F, Arslan H and Uysal I 2016 Bio-inspired Filter Banks for SS VEP-based Brain-computer Interfaces 2016 IEEE International Conference on Biomedical and Healt h Informatics (BHI) (Las Vegas, NV, USA)

work page 2016
[16]

Bakardjian H, Tanaka T and Cichocki A 2010 Optimization of SSVEP brain responses with application to eight-command Brain–Computer Interface Neuroscience letters 469 34–38 Direct information transfer rate optimisation for SSVEP-b ased BCI 11

work page 2010
[17]

Cheng M, Gao X, Gao S and Xu D 2002 Design and implementation of a b rain-computer interface with high transfer rates Biomedical Engineering, IEEE Transactions on 49 1181–1186

work page 2002
[18]

Lin Z, Zhang C, Wu W and Gao X 2007 Frequency Recognition Based on Canonical Correlation Analysis for SSVEP-based BCIs Biomedical Engineering 54

work page 2007
[19]

Friman O, Volosyak I and Graser A 2007 Multiple channel detectio n of steady-state visual evoked potentials for brain-computer interfaces Biomedical Engineering, IEEE Transactions on 54(4) 742–750

work page 2007
[20]

Zhang Y, Dong L, Zhang R, Yao D, Zhang Y and Xu P 2014 An Eﬃcien t Frequency Recognition Method Based on Likelihood Ratio Test for SSVEP-Based BCI Computational and Mathematical Methods in Medicine 2014

work page 2014
[21]

Zhang Z, Li X and Deng Z 2010 A CWT-based SSVEP classiﬁcation m ethod for brain- computer interface system 2010 International Conference on Intelligent Control and I nformation Processing pp 43–48

work page 2010
[22]

Ingel A 2017 Machine Learning in VEP-based BCI Master’s thesis University of Tartu

work page 2017
[23]

Related work T able A1

Yehia A G, Eldawlatly S and Taher M 2015 Principal component analy sis-based spectral recognition for SSVEP-based Brain-Computer Interfaces 2015 Tenth International Conference on Computer Engineering Systems (ICCES) pp 410–415 Direct information transfer rate optimisation for SSVEP-b ased BCI 12 Appendix A. Related work T able A1. Results of the related a...

work page 2015
[24]

MUSIC SVM, Gaussian kernel, one-vs-one N/A 8 Hz, 14 Hz 0.5 All N/A 63.2 1 75 2 91.1 4 94.3

work page
[25]

PSDA SVM linear Windowed sinc ﬁlter, Blackman window 8 Hz, 14 Hz, 28 Hz N/A All O1, O2, POz, Oz 65 SVM RBF 71.67 [7] BIFB No machine learning (but requires ﬁnding subject speciﬁc parameters) Bandpass ﬁlter, Hamming window 8 Hz, 14 Hz, 28 Hz 7.5 1 Oz 100 7.8 2 100 10 3 86.7 8.33 4 66.7 PSDA 10 1 66.7 9 2 66.7 15 3 60 15 4 6.7 CCA N/A 4 1 73.3 4 2 60 5 3 66...

work page
[26]

PSDA Bilinear separation N/A 8 Hz, 14 Hz 1 1 O1, O2, Oz 93 1 All ∼ 74 SVM 1 1 90 1 All ∼ 71 [4] Least square sine ﬁtting No machine learning AMUSE 8 Hz 0.5 Two subjects 8 subject speciﬁc channels 86 14 Hz 83 28 Hz 92 [15] PSDA + PCA LDA (and grid search for subject speciﬁc parameters) Bandbass ﬁlter, CAR, MAF 8 Hz, 14 Hz, 28 Hz Average over time windows u...

work page

[1] [1]

Brain-compu ter interface is a direct communication channel between the brain and some external devic e

Introduction In this work, a classiﬁcation method for steady-state visual evok ed potential (SSVEP) based brain-computer interface (BCI) is proposed. Brain-compu ter interface is a direct communication channel between the brain and some external devic e. SSVEP-based BCI uses visual stimuli, called targets, to elicit a response in the brain of t he user an...

work page

[2] [2]

SSVEP dataset for oﬄine experiments The performance of the proposed method was tested using publicly available SSVEP dataset by Bakardijan et al

Methods 2.1. SSVEP dataset for oﬄine experiments The performance of the proposed method was tested using publicly available SSVEP dataset by Bakardijan et al. ‡ which contains SSVEP data from four subjects [8]. The dataset contains EEG recordings of the subjects’ brain respons es to 8 Hz, 14 Hz and 28 Hz ﬂickering. The visual stimuli were displayed on a c...

work page

[3] [3]

Classiﬁcation method This section describes the derivation of the proposed classiﬁcation method that is based on ITR maximisation

Results 3.1. Classiﬁcation method This section describes the derivation of the proposed classiﬁcation method that is based on ITR maximisation. It follows the description of the method g iven in [14]. First, the classiﬁcation rule is described, and then a formula for est imating ITR from discrimination thresholds is derived. Finally, it is shown that the ...

work page

[4] [4]

BIFB No ML 3 N/A 8.4 88% 8.2 -

work page

[5] [5]

PSDA Bilinear separation 2 1 N/A 74% 16 -

work page

[6] [6]

This can be seen as a trade-oﬀ between accuracy and mean detection time: allowing the c lassiﬁer to not make predictions results in better accuracy but worse mean detection t ime

Discussion From the results it can be seen that allowing the classiﬁer to not make predictions when it is not conﬁdent enough results in better performance. This can be seen as a trade-oﬀ between accuracy and mean detection time: allowing the c lassiﬁer to not make predictions results in better accuracy but worse mean detection t ime. According to ITR, wh...

work page

[7] [7]

Direct information transfer rate optimisation for SSVEP-b ased BCI 10 • Combining feature extraction methods outperforms the same met hods used individually (Table 1)

Conclusion The main conclusions of the work are the following • Proposed classiﬁer outperforms previous results (Table 3). Direct information transfer rate optimisation for SSVEP-b ased BCI 10 • Combining feature extraction methods outperforms the same met hods used individually (Table 1). • Allowing the classiﬁer to not make predictions when it is not co...

work page

[8] [8]

We also acknowledge fund ing by the European Regional Development Fund through the Estonian Center of Excelle nce in IT, EXCITE

Acknowledgments The authors thank the ﬁnancial support from The Estonian Resea rch Council through the personal research grant PUT1476. We also acknowledge fund ing by the European Regional Development Fund through the Estonian Center of Excelle nce in IT, EXCITE

work page

[9] [9]

Zhu D, Bieger J, Molina G G and Aarts R M 2010 A survey of stimulation methods used in SSVEP-based BCIs Computational Intelligence and Neuroscience 2010 1–13

work page 2010

[10] [10]

Wolpaw J R, Ramoser H, McFarland D J and Pfurtscheller G 1998 EEG -Based Communication: Improved Accuracy by Response Veriﬁcation IEEE Transactions on Rehabilitation Engineering 6 326–333

work page 1998

[11] [11]

Jukiewicz M and Cysewska-Sobusiak A 2015 Implementation of Biline ar Separation algorithm as a classiﬁcation method for SSVEP-based brain-computer interfac e Measurement Automation Monitoring 61 51–53

work page 2015

[12] [12]

Karnati V B R, Verma U S G, Amerineni J S and Shah V H 2014 Freque ncy detection in Medium and High frequency SSVEP based Brain Computer Interface syste ms by scaling of sine-curve ﬁt amplitudes 2014 First International Conference on Networks Soft Compu ting (ICNSC2014) pp 300–303

work page 2014

[13] [13]

Velchev Y, Radev D and Radeva S 2016 Features Extraction Base d on Subspace Methods with Application to SSVEP BCI International Journal of Emerging Engineering Research an d Technology 4 52–58

work page 2016

[14] [14]

Anindya S F, Rachmat H H and Sutjiredjeki E 2016 A prototype of SSVEP-based BCI for home appliances control 2016 1st International Conference on Biomedical Engineeri ng (IBIOMED) pp 1–6

work page 2016

[15] [15]

Demir A F, Arslan H and Uysal I 2016 Bio-inspired Filter Banks for SS VEP-based Brain-computer Interfaces 2016 IEEE International Conference on Biomedical and Healt h Informatics (BHI) (Las Vegas, NV, USA)

work page 2016

[16] [16]

Bakardjian H, Tanaka T and Cichocki A 2010 Optimization of SSVEP brain responses with application to eight-command Brain–Computer Interface Neuroscience letters 469 34–38 Direct information transfer rate optimisation for SSVEP-b ased BCI 11

work page 2010

[17] [17]

Cheng M, Gao X, Gao S and Xu D 2002 Design and implementation of a b rain-computer interface with high transfer rates Biomedical Engineering, IEEE Transactions on 49 1181–1186

work page 2002

[18] [18]

Lin Z, Zhang C, Wu W and Gao X 2007 Frequency Recognition Based on Canonical Correlation Analysis for SSVEP-based BCIs Biomedical Engineering 54

work page 2007

[19] [19]

Friman O, Volosyak I and Graser A 2007 Multiple channel detectio n of steady-state visual evoked potentials for brain-computer interfaces Biomedical Engineering, IEEE Transactions on 54(4) 742–750

work page 2007

[20] [20]

Zhang Y, Dong L, Zhang R, Yao D, Zhang Y and Xu P 2014 An Eﬃcien t Frequency Recognition Method Based on Likelihood Ratio Test for SSVEP-Based BCI Computational and Mathematical Methods in Medicine 2014

work page 2014

[21] [21]

Zhang Z, Li X and Deng Z 2010 A CWT-based SSVEP classiﬁcation m ethod for brain- computer interface system 2010 International Conference on Intelligent Control and I nformation Processing pp 43–48

work page 2010

[22] [22]

Ingel A 2017 Machine Learning in VEP-based BCI Master’s thesis University of Tartu

work page 2017

[23] [23]

Related work T able A1

Yehia A G, Eldawlatly S and Taher M 2015 Principal component analy sis-based spectral recognition for SSVEP-based Brain-Computer Interfaces 2015 Tenth International Conference on Computer Engineering Systems (ICCES) pp 410–415 Direct information transfer rate optimisation for SSVEP-b ased BCI 12 Appendix A. Related work T able A1. Results of the related a...

work page 2015

[24] [24]

MUSIC SVM, Gaussian kernel, one-vs-one N/A 8 Hz, 14 Hz 0.5 All N/A 63.2 1 75 2 91.1 4 94.3

work page

[25] [25]

PSDA SVM linear Windowed sinc ﬁlter, Blackman window 8 Hz, 14 Hz, 28 Hz N/A All O1, O2, POz, Oz 65 SVM RBF 71.67 [7] BIFB No machine learning (but requires ﬁnding subject speciﬁc parameters) Bandpass ﬁlter, Hamming window 8 Hz, 14 Hz, 28 Hz 7.5 1 Oz 100 7.8 2 100 10 3 86.7 8.33 4 66.7 PSDA 10 1 66.7 9 2 66.7 15 3 60 15 4 6.7 CCA N/A 4 1 73.3 4 2 60 5 3 66...

work page

[26] [26]

PSDA Bilinear separation N/A 8 Hz, 14 Hz 1 1 O1, O2, Oz 93 1 All ∼ 74 SVM 1 1 90 1 All ∼ 71 [4] Least square sine ﬁtting No machine learning AMUSE 8 Hz 0.5 Two subjects 8 subject speciﬁc channels 86 14 Hz 83 28 Hz 92 [15] PSDA + PCA LDA (and grid search for subject speciﬁc parameters) Bandbass ﬁlter, CAR, MAF 8 Hz, 14 Hz, 28 Hz Average over time windows u...

work page