Seven stroke patients aged between 30 and 65, recruited from the Department of Rehabilitation Medicine of Huashan Hospital participated in our experiments (Table 1). All of them were naive to BCI and provided consent to be involved in the study. The inclusion criteria for this study were as follows: (1) unilateral motor dysfunction diagnosed by computer tomography or magnetic resonance imaging (MRI); (2) first onset stroke patient; (3) the time since stroke onset was more than 4 weeks and less than 6 months; (4) the assessment of cognitive functions: Mini-Mental State Examination score >25. The exclusion criteria were listed below: (1) unstable medical conditions; (2) severe vision problems; (3) the intervention treatment by other brain stimulations during the study period.

In the experiment, BCI intervention was performed in three sessions a week for each patient. And it lasted 1 month, with a total of 12 sessions. One session contained three runs, each run had 30 trials for each mental task (motor attempt or idle state) performed in a random order. Subjects underwent two experimental tasks. In the task of motor attempt (MA), patients were required to attempt motion of wrist extension with affected hands continually, but not to have compensatory movements. In the other task of idle state (IS), they needed to do nothing but rest (Figure 1A).

The experimental paradigm is shown in Figure 1B. In one trial, the duration time was about 11 s. The patient was asked to sit in front of a computer screen, with arms resting on a desk. A white arrow presented on the center of the screen from 0 to 3 s. The patient was instructed to keep still and rest. Then, an alternative indicator (a red square or a red rectangle) was displayed in the center of the screen, representing a task (MA or IS). After the command disappeared, the subject was required to perform the corresponding task for 5 s until the white cross disappeared. Finally, the rest interval was adopted randomly to relax.

EEG were collected using 32 channels consisting of Ag/AgCl electrode of EEG cap (actiCAP; Brain Products, Germany) according to the configuration of 10-C20 International System. A bio-signal amplifier (Brain Products) was used for acquiring the signals. The unilateral reference electrode was located in the right mastoid process, and the ground electrode was located in the forehead. The other 31 channels (FP1, FZ, F3, F7, FT9, FC5, FC1, C3, T7, TP9, CP5, CP1, PZ, P3, P7, O1, O2, P4, P8, TP10, CP6, CP2, CZ, C4, T8, FT10, FC6, FC2, F4, F8, FP2) were used for calculation. Electrode impedances were kept below 5 kΩ. The signals were amplified, digitalized with a sample rate of 200 Hz, and bandpass-filtered between 1 and 35 Hz.

BCI performance was evaluated by classification accuracy for two mental tasks. As shown in the figure of experimental paradigm, the subject conducted the mental task from 4.5 s to 9.5 s. Hence, EEG signals were extracted from [4.5 9.5] s of each single trial. And the features were filtered with the common spatial pattern (CSP) method, whose log-variance of the first and last three components were selected as output vectors. Then, a transfer learning scheme was used to improve the classification efficiency. The Logistic Regression (LR) -based classifier was used as the baseline approach. Its classification parameters w_t were calibrated using the following function,

where H, f_t, l_t and ||.||₂ denote the cross-entropy, feature vectors extracted from the EEG dataset, label vector, and 2-norm functions, respectively. Conventionally, the parameters were supposed be trained by large prior data. Transfer learning algorithms had been proposed for improving the efficiency of parameter calibration of new subjects without subject-specific data (Hossain et al., 2018). In the framework of logistic regression -based transfer learning (LRTL), a regularization term penalizing dissimilarities R_t was used for transferring the prior distribution of the existing classification parameters into the calibration of the present training of new target subjects or sessions. In this opinion, the classification parameters were calculated as follows,

and the R_t was decided by estimating the similarity between feature distribution of existing models and that of current few training data,

where μ and Σ_t were, respectively, obtained as

Here, T denoted transpose of the matrix, the functions of diag and trace were defined as the diagonal elements and the sum of the diagonal elements of a matrix. Furthermore, Kullback-Leibler (KL) divergence was added to solve the problem of weight distribution between existing models and the target model. It was supposed to give larger weights to more similar distributions and smaller weights to less similar distributions. KL divergence of two EEG sets (E₀, E₁) were represented as the following form,

where det and K denoted the determinant function and the dimension of the data, respectively. Thus, the Σ_t was updated as below,

where α and μ was computed as

Here, the divergence was calculated by averaging the KL divergences calculated for each class separately. The details of weighted LRTL (wLRTL) approach and the above other algorithms can be reviewed in Azab et al. (2019).

In this study, we discussed the transfer scheme for several subjective conditions of prior knowledge. Inter- and Intra-subject transfer learning approaches were both used to evaluate the methodological effectiveness. Inter-subject transfer calibration trained this classifier with prior experimental trials from other subjects while intra-subject transfer calibration performed this work using its own existing dataset. The target of our research was to find which kind of transfer strategies could be made to improve the online single-trial accuracy of BCI rehabilitation. The right bio-feedback (e.g., robotic arm) was able to raise the subjective confidence and patience to improve the therapeutic effect.

In this experiment, the performances of first sessions were unavailable for intra-subject transfer learning algorithms on account of no prior dataset. We used the sequential collecting of source data for transfer calibration. That meant that the target session (e.g., Session 5) would be trained by all prior collections (e.g., Session 1, Session 2, Session 3, Session 4). It was consistent with the real condition of model training. Meanwhile, the collection of source data was picked from all other subjects under the inter-subject condition. Specifically, the dataset of each subject which obtained the best performance in all sessions was used for transfer learning. Moreover, 5-fold cross validation was conducted for each approach. All 31 channels of the EEG data were selected for pattern classification. The 45 trials of MA and IS tasks (45 trials per class) were randomly divided into five sets. Four sets were used to train the classifier and the other set was tested to evaluate the performance.

In our study, the precision of pattern recognition was considered as the most important index for BCI rehabilitation. Figure 2 lists the classification results of the above intra-subject transfer learning approaches (LRTL, wLRTL), as well as the baseline approach (LR). The average classification accuracies of all patients were higher than 60% for three different algorithms, except for P6. And it was indicated that the pattern of motor attempt could be distinctive from that of an idle state without motor attempt. However, a paired t-test with Bonferroni correction showed that no discriminatory differences were presented between transfer learning approaches and the baseline approach (LR vs. LRTL: p = 0.0277; LR vs. wLRTL: p = 0.0613; LRTL vs. wLRTL: p = 0.6085). This result suggests that transfer calibration did not significantly improve the BCI performance.

Furthermore, we analyzed the performance of low-precision (≤ 60%) sessions for all patients (Table 2). Paired t-test with Bonferroni correction showed that the classification results of transfer learning approaches were significantly greater than that of the baseline approach (LR vs. LRTL: p = 0.0001; LR vs. wLRTL: p = 0.0001; LRTL vs. wLRTL: p = 0.0563). It was meaningfully revealed that transfer calibration could improve the BCI performance induced by poor model training.

Similarly, Table 3 lists the classification results of all three algorithms under the inter-subject condition. Paired t-test with Bonferroni correction showed that no discriminative differences were presented between transfer learning approaches and the baseline approach (LR vs. LRTL: p = 0.0488; LR vs. wLRTL: p = 0.1207; LRTL vs. wLRTL: p = 0.1744). This result suggested that the non-significance was consistent with those under the intra-subject condition.

Additionally, low-precision sessions were extracted for further analysis (Figure 3). In this case, the statistical analysis indicated that the accuracies of transfer learning approaches were significantly higher than that of the baseline approach (LR vs. LRTL: p = 0.0001; LR vs. wLRTL: p = 0.0001; LRTL vs. wLRTL: p = 0.0210). It was revealed that transfer calibration improved the BCI performance under the inter-subject condition.

As we know, the critical issue of BCI rehabilitation revolves around how to promote the biofeedback effect for active intervention (Ko et al., 2019). It was positive for rehabilitation outcomes, enhancing cortical activity for neural recovery, and increasing confidences of voluntary training (Zhang et al., 2020). Hence, the biofeedback of low-precision trials was negative for patients. Furthermore, most of the sessions used by the baseline classifier achieved the effective performance (>60%) for each patient, except for S6. It was indicated that very few non-effective results of experimental sessions were unreliable for evaluating this neural treatment. Improved performance of transfer calibration could eliminate the confusion caused by the precision fluctuation.

Moreover, we presented the comparison of transfer learning algorithms between inter-subject (IERS) and intra-subject (IRAS) conditions (Figure 4). Statistical analysis was used to compare the classification accuracies among transfer conditions (LRTL: IERS vs. IRAS: p = 0.3138; wLRTL: IERS vs. IRAS: p = 0.3501;). This result suggested that both of them were reliable for improving low-precision performance of SMR-BCI tasks. However, the advantage of inter-subject transfer calibration was employed for new subjects without training.

The fluctuation of classification for LR resulted from the differences of brain activity between consecutive sessions. It was deduced that changes of cerebral activities were caused by neural self-rehabilitation for patients. As a result, BCI-based intervention was effective for stroke rehabilitation to some extent. Nevertheless, the methodology needs to be further clarified for high-efficiency treatment.

The efficiency of transfer learning has been verified by motor imagery -based BCI tasks for healthy people. However, cerebral impairment of stroke patients would influence the training effect due to the weak neural activities. Therefore, our classification result was lower than those of the above state-of-art BCI systems (Azab et al., 2019, 2020; He and Wu, 2019). Nevertheless, the improvement of low-precision performance was conductive to treatment for the impatient. Compared to the feedback of the random level, the subject was subjectively motivated by positive feedback of right detection. It is meaningful for long-term continuous rehabilitation.

In our study, an available scheme of transfer calibration was proposed for model selection in the online task of SMR-BCI rehabilitation (Figure 5). We summarized several rules as stated below: (1) Instant self-training was necessary for new subjects. The classification model based on current dataset was reliable for BCI tasks. (2) If the patient was frustrated by tedious training, intra- or inter-subject transfer calibration could be used to reduce the calibration time. (3) Furthermore, we could train another model when sufficient trials were finished in the task. If the precision of current model was superior to prior transfer model, the alternative could be automatically performed by our control system. (4) If the model based on the current training dataset performed poorly on the experiment, intra- or inter-subject transfer calibration was worth trying in order to replace the under-performing model. (5) For transfer calibration, the volume of EEG data was a crucial factor for model selection between intra-subject and inter-subject conditions. Sufficient training data was an essential precondition for transfer calibration. Specifically, the only option for new subjects without prior experience was inter-subject transfer calibration.

This scheme of transfer calibration was feasible for improving the poor performance of SMR-BCI recognition. And it also reduced the time consumption of model calibration. It was inferred that this scheme was suitable for other transfer learning algorithms.

In this study, some issues should be noted and considered in our future work. First, the scheme of transfer calibration needed to be verified for a large amount of stroke patients. It will be addressed in future studies. Second, these patients performed these experiments for 3 months. BCI training over longer time periods should be observed to evaluate performances of the patients in different stages of post-stroke time. Moreover, the number of electrodes was supposed to be reduced by data analysis. It could reduce the time consumption of BCI rehabilitation. Thus, future studies should be conducted to solve these problems to improve the performance of online SMR-BCI rehabilitation.