Separable EEG Features Induced by Timing Prediction for Active Brain-Computer Interfaces

Brain–computer interfaces (BCI) have witnessed a rapid development in recent years. However, the active BCI paradigm is still underdeveloped with a lack of variety. It is imperative to adapt more voluntary mental activities for the active BCI control, which can induce separable electroencephalography (EEG) features. This study aims to demonstrate the brain function of timing prediction, i.e., the expectation of upcoming time intervals, is accessible for BCIs. Eighteen subjects were selected for this study. They were trained to have a precise idea of two sub-second time intervals, i.e., 400 ms and 600 ms, and were asked to measure a time interval of either 400 ms or 600 ms in mind after a cue onset. The EEG features induced by timing prediction were analyzed and classified using the combined discriminative canonical pattern matching and common spatial pattern. It was found that the ERPs in low-frequency (0~4 Hz) and energy in high-frequency (20~60 Hz) were separable for distinct timing predictions. The accuracy reached the highest of 93.75% with an average of 76.45% for the classification of 400 vs. 600 ms timing. This study first demonstrates that the cognitive EEG features induced by timing prediction are detectable and separable, which is feasible to be used in active BCIs controls and can broaden the category of BCIs.

Obviously, with the increasement of training days, subjects had a significant ascending trend for response accuracy (F(2, 88) =286.124; P<0.001) but a significant descending trend for reaction time (F(2, 88) =116.453; P<0.001). To be specific, (sFig.1b) the accuracy of the third day (92.36%) was significantly higher than those of the second (83.39%, P<0.001 after Bonferroni correction) and the first day (62.58%, P<0.001 after Bonferroni correction), and the accuracy of second day significantly higher than that of the first day (P<0.001 after Bonferroni correction). The reaction time was also significantly reduced by training (F(2, 88) =287.121; P<0.001). Specifically, the reaction time was the shortest for the third day (506.15ms), medium for the second (638.46ms) and longest for the first day (1009.17ms), as displayed in sFig.1b. It demonstrates the effectiveness of trainings for the subjects on the constructing the idea about the three timing intervals in their minds.
Notably, as the ability of perceiving timing intervals varied with subjects, so the training blocks for each subject were different. In each training day, the subjects with better time perception, only did about 5 training blocks; whereas the ones who was insensitive to perceive the timing intervals, needed to train for 10~20 blocks. For this reason, the current study mainly calculated the averaged behavioral result of each training day. As to the sentence "at least three training blocks successively", it was monitored by the experimenter during training.

S2: lateralization potentials for measuring the influence of button-press
Allowing for the subjects needed to make judgement by pressing button, it remains unclear whether it could influence the neural signatures of timing prediction or not. Therefore, we calculated the lateralization potentials to measure the influence of button-press on the ERP profiles. After aligning the data to the button-press moment (zero point in sFigure.2(a)), the lateralization potential of typical motor-related electrodes (C3, C4) was calculated as follows: Where represents lateralization potential, represents the amplitude in C3 electrode when subjects pressed button using right hand (contralateral), while represents amplitude in C3 electrode when subjects pressed button using left hand (ipsilateral); so did C4. When the lateralization potential is zero, it means movement does not have an influence on ERP profiles, whereas the larger its value is, more obvious the its influence is.
sFigure.2(a) showed the lateralization potentials induced by button-press. Obviously, in T400 and T600 conditions, the lateralization potentials emerged at about 200ms before buttonpress, while that of NT emerged earlier. It is reasonable for the NT to result in earlier lateralization potential, as the subjects constantly pressed the button by only one hand in whole experimental block. sFigure 2(b) and 2(c) were the amplitude topographies covering -250~-200ms and -50~0ms relative to the button-press moment. It is evident that the former period was rarely influenced by button-press, whereas the later period was influenced. These observations indicated, in this study, the influence of movement was mainly within 200ms before button-press. Thus, during the period that we selected for classification, ERP separations between T400 and T600 is mainly attributed to timing prediction, rather than button-press.

S3: combined DCPM and CSP method is more suitable for current EEG features
In future studies, we may be able to find better classification methods. However, we think the current method, i.e., the combined DCPM and CSP method, is more suitable for the EEG features in this study. In supplementary materials, we compared the current methods with part of other machine learning methods in four aspects. Firstly, we compared the effects of classifier.
In the current study, after the feature extraction, Fisher discriminative analysis (FDA) is used for distinguishing the T400 and T600 signals, we argue that the effect of FDA is similar to other classifiers, such as the support vector machine (SVM). In supplementary materials sTable1, the classification accuracies of DC+PM, DC+FDA, DC+SVM were compared. It should be noted that the DCPM method contains both feature extraction (DSP filter and CCA method, I.e., DC) and classification (Pattern Matching, PM) procedure. After the 'DC' feature extraction, there were three methods used for classification, i.e., PM, FDA and SVM.
sTable 1 classification accuracy based on 0~3Hz data. Thirdly, there were two kinds of combination methods, i.e., combination in decision level and in feature (for details about the two methods, please see reference 'enhance decoding of pre-movement EEG patterns for brain-computer interfaces'). As shown in sTable4, the two methods had almost similar performance.

Subject
Fourthly, some deep learning algorithm may have better performance, but the data set of current data was limited (40 trials at most), which obstructs the use of these algorithms. In sum, the current classification method is more suitable for the current EEG features.