Adaptive CSP for user independence in MI-BCI paradigm for upper limb stroke rehabilitation

Have your doctor explain how these calculations will get you 100% recovered.  You'll have to go to the link to see them properly.

Adaptive CSP for user independence in MI-BCI paradigm for upper limb stroke rehabilitation

Ana P. Costa, Jakob S. Møller, Helle K. Iversen and Sadasivan PuthusserypadyAbstract—A 3-class motor imagery (MI) Brain-ComputerInterface (BCI) system, that implements subject adaptationwith short to non-existing calibration sessions is proposed. Theproposed adaptive common spatial patterns (ACSP) algorithmwas tested on two datasets (an open source data set (4-class MI),and an in-house data set (3-class MI)). Results show that whenlong calibration data is available, the ACSP performs only slightlybetter (4%) than the CSP, but for short calibration sessions, theACSP significantly improved the performance (up to 4-fold). Aninvestigation into class separability of the in-house data set wasperformed and was concluded that the “Pinch”movement wasmore easily discriminated than “Grasp” and “Elbow Flexion”.The proposed paradigm proved feasible and provided insights tohelp choose the motor tasks leading to best results in potentialreal-life applications. The ACSP enabled a successful semi userindependent scenario and showed potential to be a tool towardsan improved, personalized stroke rehabilitation protocol.Index Terms—Brain-computer interface (BCI), Stroke rehabil-itation, Sensorimotor rhythms (SMR), Adaptive Common SpatialPatterns (ACSP)I. 

INTRODUCTION 

Brain-Computer Interface (BCI) technology allows for brainsignals to be recorded and translated into output commands,which can be used in various applications. In electroencephalogram (EEG) setups, sensorimotor rhythms (SMRs) are someof the signals of interest which can be measured. SMRs are tuned by motor intentions, such as motor imagery (MI), and are characterized by a modulation of the amplitudes of the measured electrical potentials.One area where MI-BCI systems have real-world applications is in neuro-rehabilitation, namely in stroke cases. Current stroke rehabilitation therapies present some limitations [1]–[3]and different enhancing strategies are emerging, of which BCIs are a promising one [4]. Rehabilitative BCIs aim at exploiting brain plasticity to improve motor recovery in patients. MI is used in most studies for this purpose, as it is hypothesized that it promotes neuroplasticity-related repair of the damaged brain areas [5]. The basis of MI-BCI systems for stroke rehabilitation has been laid by studies reporting an increasein motor cortex excitability as well as topographical changesafter training [6]. Preliminary results such as [7], [8] indicate the feasibility of incorporating BCI in post-stroke hand rehabilitation. Nevertheless, more large, randomized clinical trialsare necessary to confirm the advantages and reliability of themethod.Another issue to consider, is that individual stroke characteristics lead to different consequent neuroplastic changes during recovery, which indicates that an ideal system should be tailored for each patient [9]. This is related to one ofthe disadvantages of many BCI systems, which is subject-dependence: systems require data from a long training session for each subject, where no feedback is given to the user. This is impractical and particularly undesirable in the context ofstroke rehabilitation, where it is important that the patients start receiving feedback as soon as possible.A. Adaptive Spatial FiltersA commonly used strategy for source localization in MI-BCI systems is the common spatial pattern (CSP) filters.However, it presents some setbacks, namely that (i) it requireslarge data to avoid overfitting and generate robust projectingvectors, while being quite sensitive to outliers and (ii) it istypically subject-dependent and has no ability to adapt tothe non-stationarities that are characteristic of EEG signals[10]. These characteristics imply that long subject dependent calibration sessions are needed for computing the filter coefficients. Therefore, variations of the algorithm are needed to solve these problems and improve its performance. There aretwo ways to incorporate new data in order to handle changes that occur between distinct EEGs: block-wise [10], [11] and sample-wise. Here, a Recursive Least Squares (RLS) approach was implemented for sample-wise adaptation of the CSP filter,similar to the method used by [12] for the axDAWN filter.II. 

 MATERIALS AND METHODS

 A. Signal Processing1) The CSP Filter:LetXj∈RCn×Nbe thejthEEGtrial, whereNis the number of samples/trial and Cn is the number of channels. Then, the filtered trialZj,CSP∈RCn×N=W Xj, whereW∈RCn×Cnis the matrixparameterizing the signal decomposition. Here, we denote each column (wi,i=1,2,...,Cn) ofWas a spatial filter, andeach column ofW−1as a spatial pattern. The normalizedcovariance matrix for classkis defined as:Ck=1TnTn∑j=1Xj(k)XTj(k)trace{Xj(k)XTj(k)},(1)whereTnis the number of trials,Xj(k)is the jth trial belonging to classk∈[1,K]andK= 2for the binary classification case, which we will use, for simplicity, to explain420978-1-7281-1295-4/18/$31.00 ©2018 IEEEGlobalSIP 2018
the algorithm. In [13], the CSP is described with a discrimina-tive view, which was useful to derive the ACSP. Considering Cd=C1−C2andCc=C1+C2as the discriminative activity (Cd),i.e.the band-power modulation between the two classes, and the common activity (Cc), the solution to the following maximization problem can be achieved by solving the GED problem:argmaxWWTCdWWTCcW.(2)A One-Versus-All (OVA) multiclass version of the CSP was used so that the algorithm can be used to distinguish between nmore than two classes [14]. Finally, to get a system that is trainable on a small amount of data, the CSP was regularized with Diagonal Loading (DLCSP algorithm) [15].2) Adaptive Spatial Filter (ACSP):A sample-wise adaptive approach to the CSP algorithm based on the RLS method forGED is introduced here [16]. A training set is used to initializethe CSP matrix. Expanding Eq.(2), it can be deducted that:argmaxWWTCdWWTCcW= argmaxWWTC1WWTCcW.(3)BothC1andCcrepresent full normalized covariancematrices of stationary signals with zero mean. Rearranging the solution to the GED problem as in [16], we obtain the basis for the iterative algorithm:W=WTCcWWTC1WC−1cC1W.(4)A temporal discrete variablen is now introduced and an estimate of the primal eigenvectorw1(n)is computed as:ˆw1(n) =wT1(n′)Cc(n)w1(n′)wT1(n′)C1(n)w1(n′)C1c(n)−1C11(n)w1(n′)(5)where n′=n−1. Some comments about Eq.(5):1)Iterative computation of class covariance matrices:Here,we have not prioritized an asynchronous, self-paced systemwhich results in an advantage for the development of the ACSP, specifically in this step, because we always know the true label of each samplex(n). Therefore, we can iteratively update the normalizedC1(n)only whenx(n)∈class1, whileCc(n)is always updated:C1(n) =C1(n′) +x1(n)xT1(n)trace{x1(n)xT1(n)}andCc(n) =Cc(n′) +x(n)xT(n)trace{x(n)xT(n)},(6)wherex(n)is any data sample taken at timenandx1(n)represents the data belonging only to class 1.2)Iterative computation of the inverse ofCc(n):This stepin Eq.(5) would imply a very high computational effort whichis not feasible for an online application. Therefore, as in [12],[16], the Sherman-Morrison-Woodbury formula is used for theiterative update ofC−1c(n)[17]:C−1c(n) =C−1c(n′)−C−1c(n′)x(n)xT(n)C−1c(n′)1 +xT(n)C−1c(n′)x(n).(7)It is advantageous to use Eq.(7) since only C−1c(n′)needs to be stored and only simple matrix operations are required for each iteration. Finally, a deflation technique [16] is used to iteratively estimate the remaining eigenvectors (wi’s).Ci1(n) =[I−Ci−11(n)wi−1(n)wTi−1(n)wTi−1(n)Ci−11(n)wi−1(n)]Ci−11(n),Cic(n) =Ci−11(n).(8)To computew′is, Eq.(5) is used with the respectiveCi1(n)andCic(n)computed as in Eq.(8). As in [12], thew′is are normalized for numerical reasons.wi(n) =ˆwi(n)[ˆwTi(n)Cic(n)ˆwi(n)]12.(9)After all the spatial filters of trialnhave been computed,the first and lastmvectors are stored to classify trial n+ 1.3) Classification Algorithm:Discriminant Analysis (DA)was the chosen method to perform the classification, due to itslow computational complexity and comparable performances to more complex approaches [12], [18], [19]. To overcome some of the limitations of DA, Friedman’s regularized version of DA [20] (RDA) was implemented [21].B. Dataset Description and Experimental Design1) Dataset 1: 4-class MI of different body parts:This dataset belongs to the BCI competition IV [22] (dataset 2a)and comprises of the EEG recordings on 9 subjects of four classes of MI from distinct body parts (left and right hand,feet and tongue). For each subject, two sessions of 288 trialswere recorded, namely a calibration session without feedbackand an evaluation session with feedback.2) Dataset 2: 3-class MI of single upper limb:This dataset was recorded in our laboratory and was obtained in two differ-ent days: a short calibration session recorded without feedback and a longer session with feedback. The three motor tasks (Fig.1) were performed with the right arm. A goal-oriented visualinterface was implemented, as it has been proven to improve classification results [18]. For the calibration day, each sessionof MI consisted of 6 runs of 18 trials. Each trial started with a warning sound and a fixation cross, for the subject to mentally prepare for the task (2 seconds). Then, it was replaced by a visual cue indicating which of the three motor tasks the subjectshould imagine (4 seconds). Finally, the screen became blankand the subject could rest (2 seconds). In the online session,a vertical green bar positioned on the right side of each cuedisplayed real-time feedback (the bar grew from bottom totop, one fragment at a time for correct classifications). Theuser started receiving feedback after 1 second.The recordings were done on 14 healthy subjects agedbetween 20-31. The subjects were sitting comfortably in achair placed approximately one meter from the screen display-ing the visual interface. During the experiment, the subjects421(a)(b)(c)Fig. 1. Visual interface of the BCI system. (a) Palmar grasp (Class 1) -engaging all fingers and palm to hold an imaginary object between them. (b)Pinch (Class 2) - collecting the fingertips of the thumb, index and middlefinger. (c) Elbow Flexion (Class 3) - flexing the elbow while maintainingthe wrist aligned with the arm, with the thumb directed upwards/towards thesubject as the forearm is lifted.placed their right hand comfortably on the table in frontof them and kept all movements to a minimum. 16 activeAg/AgCl electrodes were used spanning the motor cortexarea. All procedures involving human subjects were performedin accordance to the ethical standards of the 1964 Helsinkideclaration and of the national research committee.C. Data Analysis SetupBefore analysis, the data was band-pass (7-30 Hz) filteredusing a4thorder zero-phase Butterworth filter. Two distinctstrategies were used to asses the performance of the ACSPfilter on dataset 1: (1)User dependent strategy:one CSP filterand RDA classifier were trained for each subject using all thedata from the calibration session and tested on the evaluationdata, and (2)Semi user independent strategy:shorter calibra-tion sessions were used to initialize the feature extraction andclassification parameters. This allows for potential customiza-tion of the BCI system for the individual needs of each patient.After testing the ACSP on dataset 1, it was used on dataset2, where the training size was short and determined based onthe previous results.III. 

RESULTS AND DISCUSSION 

A. Dataset 1: 4-class MI of different body parts1) User dependent strategy:An investigation of the con-vergence of the adaptive filter was made prior to the analysis of the classification performances. In Table I, the classification performances of the CSP and ACSP in the unseen evaluation data are displayed and compared to the results of the winningalgorithm (filter bank CSP (FBCSP)) of the BCI competition[14], with the best performance for each subject highlightedin bold. While a one-sided paired t-test indicated that thedifference between the performance of the FBCSP and thatof the ACSP algorithm was non-significant (p-value of 0.141at a confidence level ofα= 0.05), the first still outperformedthe latter in all subjects except three. A similar paired t-test revealed that there is no significant difference between the CSP and the ACSP (p= 0.294). These results indicate that there is little advantage in using the ACSP algorithmwhen sufficient training data is available. Finally, the scalptopographies were analyzed and it was concluded that the ACSP lead to physiologically significant patterns similar tothe ones obtained by the regular CSP algorithm.TABLE ICLASSIFICATIONPERFORMANCE(ASMAXIMUMKAPPAVALUE)OFACSP, CSPAND THEWINNER(FBCSP)OF THEBCI COMPETITIONIVIN THEUNSEENEVALUATIONDATA OFDATASET1.SubjectsCSPACSPFBCSP [14]10.6770.6830.67620.3630.2310.41730.6020.6770.74540.4650.3770.48150.2460.3300.39860.2430.3660.27370.6120.5680.77380.7490.7040.75590.5650.7710.606Mean0.5020.5230.569Median0.5650.5680.6062) Semi user independent strategy:In this approach, thenumber of trials per class was made to vary from 15 to 65 insteps of 15 and the resulting kappa values corresponding tothe classification performances on the evaluation dataset were calculated for each case. The trials chosen for each class weretaken randomly. The result is presented in Fig. 2. It is clear2030405060700.250.350.450.50.55Kappa valueAdaptive CSPfixed CSPFig. 2.Evolution of classification performance on evaluation dataset for different sizes on the training session (online simulation). Samples of growing size from the calibration dataset were used to train the CSP, for each subject, and evaluated on the corresponding evaluation dataset, using the same procedure as in section III-A1. The average maximum kappa value of all subjects is used as evaluation performance.that the performance of the CSP decreases significantly withsmaller training sizes. The ACSP, however, results in kappavalues similar to the final one already from a very small set.While the difference in performance kappa between algorithmsfor a training size of 72 is only 0.021, for a training size of 35the difference is 0.115. This represents a 1.28-fold increase.Based on this analysis, the training sizes for dataset 2 were chosen to be 36.B. Dataset 2: 3-class MI of single upper limbA similar convergence analysis was made for dataset 2and it indicated convergence around 17 trials per class. In Table II, the classification performances obtained during the online feedback session in terms of average maximum kappavalues for the MI are summarized. We conclude that the422TABLE IIPERFORMANCES FROM THEMI ONLINESESSION OFDATASET2WITHREALTIMEFEEDBACKGIVEN ASAVERAGEMAXIMUMKAPPAVALUE.Subjects124567891011121314MeanMedianDLCSPFixed0.100.080.050.190.050.100.080.050.210.050.300.130.080.110.08Adaptive0.360.490.520.460.490.650.500.710.6510.480.330.330.220.470.49ACSP resulted in a 4-fold increase in performance of theDLCSP, which is a significant result (p= 5×10−5). Theperformance of the ACSP is comparable with the literaturefor similar problems [18], [23]. Finally, an investigation intoclass separability was made, by extracting and combining thebinary confusion matrices for each class (Table III). The resultssuggest that class 2 (Pinch) was the easiest to discriminate, andclass 3 (Elbow Flexion) the hardest. 

TABLE 

IIICLASSIFICATION PERFORMANCES FOR PAIRWISE CLASS DISCRIMINATION,FOR THE ONLINE FEEDBACK SESSION.Class combination1 vs 21 vs 32 vs 3Kappa value0.6320.5160.579IV. 

 CONCLUSION 

The feasibility of a 3-class MI-BCI paradigm which could be employed for enhancement to the current stroke rehabilitation therapies has been studied.(Who the fuck cares about study? We need solutions.) The RLS-based ACSP seems to   overcome one of the main disadvantages of the CSPfilter, allowing for personalized training programs based on short calibration sessions. The ACSP provided only slightly better results than the CSP when there was plenty training data, but performed up to 4 times better when not. The classification performances are lower in the second dataset,confirming that it is a harder task to distinguish between motor tasks performed by the same limb. An investigation onthe separability of the chosen motor tasks indicates that the“Pinch” movement was the easiest to discriminate, which can suggest a direction for class choice in future similar studies.The system here implemented could be(Once again followup needed) a step towards a potential application of MI-BCI technology for enhancementof current post-stroke neuro-rehabilitation. Overall, there isstill room for improvement towards practical applicability,such as channel reduction and development of an unsupervisedversion of the algorithm. Large and randomized clinical trialsare also necessary to confirm the advantages and reliability of the method.