gms | German Medical Science

Common spatial pattern (CSP) is an algorithm which decomposes a signal into spatial patterns that are extracted from multiple classes. These are used to calculate spatial filters that maximize the ratio of the variance of one class to another [1], [2]. It can be utilized to decompose an electroencephalography (EEG) signal into components that separate different classes [1], [2], which in turn can be used to control a brain-computer interface (BCI). BCIs can be used to control different types of hardware and software, such as wheel chairs [3] or an avatar in a virtual reality environment [4]. Integration of BCIs into the rehabilitation therapy of stroke patients is an ongoing research subject [5], [6].

Many variants of CSP exist and are still subject to research [7], [8], [9]. As CSP is used to create signal-specific filters; an interesting characteristic of CSP is the ability to reveal information about the origin of the created filter. This can be utilized to select important features or electrodes for a specific subject [10], [11]. Wang et al. calculate the event-related desynchronization and readiness potential using all electrodes and select only those electrodes that have maximum value based on the computed CSPs [10]. Another approach is omitting channels where the average of the CSP mixing matrix coefficients is less than a certain threshold and evaluate the remaining signals with the help of a neural network [11]. However, the authors did not state the impact on the electrode reduction and did not compare it with other approaches.

In this study, our objective is to identify improved electrode combinations for classifying an imaginary movement (IM) task through the use of CSP. We aim to compare these combinations with CSP calculations incorporating all available electrodes.

In this work two different approaches are proposed:

• Approach 1: Check every possible combination on one dataset, take the best combinations and try them on the other datasets (to see if some generalization is possible).
• Approach 2: Infer possible combinations for each subject in each dataset using the patterns from the base CSP calculation.

Another approach would be to try every possible combination for each subject in each dataset and choose the best one. Due to the high computational expense this approach is unfeasible and therefore was discarded.

To acquire the needed data we utilize the library “Mother of all BCI Benchmarks” (MOABB) [12] as it contains interfaces to different freely available BCI datasets. Only datasets that fulfill the following conditions were considered for the evaluation:

• All sessions in the dataset are recorded with at least 90 Hz sampling rate.
• The paradigm of the dataset contains at least a left-hand imagination and a right-hand imagination class.
• The electrodes used for data acquisition contain at least the electrodes used in [13] (see Figure 1 [Fig. 1]).

Four datasets meet these conditions: Yi2014 with n=10 [14], BCI Competition IV Dataset 2a (BNCI2014) with n=9 and 2 sessions per subject [15], Cho2017 with n=49 [16] and PhysioNet with n=109 [17].

We use the EEG data from the datasets with the left-hand imagination and right-hand imagination classes (which consist of single finger or full hand movements). Subsequently we utilized a simple pipeline as shown in Figure 2 [Fig. 2] containing: 1) a preprocessing step, where the data is reduced to the specified electrodes and bandpass filtered between 5 and 45 Hz, 2) predefined cross-validation splits, 3) calculation of the CSP filter for each split and filtering the signal with it using a CSP implementation from MNE-Python [18], 4) a classification step with linear discriminant analysis (LDA) and 5) the combination of the accuracies to a mean accuracy.

Baseline: We perform a baseline calculation with the pipeline utilizing all electrodes for each subject from the dataset (BFull) as well as the pipeline with 16 electrodes that are positioned as shown in Figure 1 [Fig. 1] (B16).

Approach 1: We consider only the 16 electrodes leading to an amount of possible combinations of . We choose to only use combinations that contain at least 8 electrodes to have enough features for LDA after the CSP filter, which results in combinations. Further we decided to use Yi2014 as base dataset as it is fast to calculate for each combination. We run the pipeline for all combinations on each subject of Yi2014. The percentage change between the 16 electrode baseline accuracy (of Yi2014) and each combination’s accuracy is calculated afterwards. Last, we assess whether the expectation µ is significant better over all subjects for each combination using a t-test with the following hypothesis:

(1) H₀:µ≤0, H_a:µ>0 and α=5%

Significant combinations are used for processing the other datasets.

Approach 2: We developed an algorithm to infer more useful electrodes by using the CSP calculation from the baselines. Therefore, we use the pattern p from the first 8 CSP components [the amount of components calculated with CSP is the amount of electrodes, but we only consider the first 8 as the latter are less significant] and select the electrodes that fulfill the following condition:

(2) |e_p–|>θ·σ(p)

where e_p is the value of the electrode in the pattern p, is the mean of all electrode values of the pattern, θ is a predefined threshold and σ(p) is the standard deviation of all electrode values of the pattern. In our experiments we set θ=1.5. Figure 3 [Fig. 3] shows a visualization of the algorithm.

We calculate the electrodes for each cross validation split. The found electrode combinations (containing more than 2 electrodes) are then used to calculate the accuracy for the subject and the combination with the highest accuracy is furthermore compared against the baselines. The complete pipeline is visualized in Figure 4 [Fig. 4].

To evaluate our approaches we perform Bayesian A/B tests [19], [20]. We compare different variants against the full electrode baseline (BFull). The used variants are:

• The 16 electrode baseline (B16)
• The best electrode combinations from approach 1 (A1)
• The per subject best electrode combinations using our algorithm with 16 electrode base (Algo16)
• The per subject best electrode combinations using our algorithm with full electrode base (AlgoFull)
• The per subject best electrode combinations of the combinations from approach 1 (PSA1)
• The per subject best electrode combinations of the combinations from approach 1 and our algorithm (both bases) (Comb)

We use the beta distribution for our prior and posterior distribution as we have dichotomic data.

Further we assume a priori that all variants perform better than average so we set α=6 and β=4. However the prior has a marginal effect due to the amount of samples provided. Afterwards we perform a Monte-Carlo simulation with n=10,000 for our variants and calculate the fraction of each variant against the full electrode baseline. By calculating the mean of these fractions we get a percentual improvement against BFull.

After testing all combinations on Yi2014, 26 remain as significant over all 10 subjects from Yi2014. Applying those combinations separately on the other datasets did not lead to improvements of the accuracy compared to BFull (see Table 1 [Tab. 1] and A1 in Table 2 [Tab. 2]), whereas using only the best of all 26 combinations for each subject individually increases the accuracy significantly (see PSA1 in Table 2 [Tab. 2]). Using the proposed algorithm improves accuracy compared to BFull, but it is not as effective as PSA1 (see Algo16 and AlgoFull in Table 2 [Tab. 2]). Interestingly combining the different approaches leads to an improvement of more than 10% (see Comb in Table 2 [Tab. 2]).

These findings are also visualized in Figure 5 [Fig. 5], which shows the computed beta distributions of where the samples are taken from for the percentual improvement calculation.

As BNCI2014 contains 2 sessions per subject, we test whether a good performing combination in one session also improves the accuracy for the other session. Figure 6 [Fig. 6] shows the result for subject 1 and 3. It is obvious that a combination which works well for one session may not perform equally well in another session.

Our results show that using only specific electrodes has the potential to increase the accuracy of an IM task compared to using all electrodes (as also shown in [21]). Reducing the amount to a predefined number of electrodes (from all to 16) does not lead to better results. We also showed that even for the same subject the calculated electrode combination cannot be transferred to another session. In summary, it can be seen that it is important to calculate a well-performing electrode combination for each session and subject so that the overall amount of electrodes applied to the subject cannot be reduced, but the evaluation speed and accuracy of an IM task can. The proposed algorithm for determining good performing electrode combinations can be used even in a clinical setting, as it is fast to calculate and EEG tasks usually need calibration anyway. Furthermore, we proposed 26 electrode combinations that can be considered additionally for better performance. In summary, sophisticated selection of the appropriate electrode combination can enhance the accuracy of an IM task. In our research we achieved an improvement of over 10%.

To further improve performance the algorithm to detect electrode combinations could be revised or changed as there are subjects where it performs worse. It was assumed that this is due to a low amount of channels. However, there is no correlation between the number of channels and the percentage of change. Another step could be to determine another set of well-performing electrode combinations using the brute-force calculation (approach 1) on another dataset and compare it to the shown results.

The source code for this research is available at https://gitlab.com/domwet/csp-research.git.

The underlying datasets can be accessed as follows:

• Yi2014 [14]: https://doi.org/10.7910/DVN/27306 [22]
• BNCI2014 [15]: https://doi.org/10.21227/katb-zv89 [23]
• Cho2017 [16]: https://doi.org/10.5524/100295 [24]
• PhysioNet [17]: https://doi.org/10.13026/C28G6P [25]

The authors declare that they have no competing interests.