1 Introduction
-
A covariate shift-adaptation model is introduced to address the effects of non-stationarity in the EEG signals.
-
An EWMA-based CSD test is applied to detect the non-stationary changes in the principal component analysis (PCA)-based features of the motor imagery-based brain responses.
-
Third, the proposed model updates its classification decision boundary online without making any a priori assumption about the distribution for the upcoming test data.
2 Methods
2.1 Adaptive learning problem formulation
2.2 Algorithm overview
-
CSD: CSD test monitors the stationarity of \(x_i \), disregarding their supervised labels.
-
\(\mathcal {F}:\) The pattern classifier \(\mathcal {F}\) is used to classify the input samples.
-
KB: The current knowledge base (KB) updated on each CSD.
2.3 Covariate shift-detection (CSD)
2.4 Covariate shift-validation
2.5 Covariate shift-adaptation
2.5.1 Transductive learning with CSD (TLCSD)
2.5.2 Adaptive learning with CSD (ALCSD)
3 Application to brain–computer interface
3.1 Data description
3.1.1 BCI Competition IV dataset 2A
3.1.2 BCI competition IV dataset 2B
3.2 Data processing and feature extraction
3.2.1 Temporal filtering
3.2.2 Spatial filtering
3.3 Covariate shift-detection (CSD)
3.4 Experimental setup and classification evaluation metrics
4 Results
4.1 Results for dataset 2A
Subject | Lambda | Shift-detected | Shift-validated |
---|---|---|---|
A01 | 0.10 | 7.64 | 2.78 |
A02 | 0.80 | 7.64 | 6.25 |
A03 | 1 | 4.86 | 1.39 |
A04 | 1 | 7.64 | 4.17 |
A05 | 0.30 | 10.42 | 4.17 |
A06 | 0.10 | 9.72 | 3.47 |
A07 | 0.10 | 8.33 | 6.25 |
A08 | 0.20 | 7.64 | 3.47 |
A09 | 0.50 | 6.94 | 2.78 |
10-CV | Baseline | SSL |
\(\text {TLCSD}_1 \)
|
\(\text {TLCSD}_2 \)
| ALCSD | UB | |
---|---|---|---|---|---|---|---|
Tr | Eval | Eval | Eval | Eval | Eval | Eval | |
A01 | 85.71 | 89.58 | 79.17 | 90.28 | 90.28 | 90.28 | 90.28 |
A02 | 75.71 | 53.47 | 54.17 | 57.64 | 57.64 | 54.17 | 58.33 |
A03 | 92.86 | 92.36 | 93.06 | 93.06 | 95.14 | 93.75 | 97.22 |
A04 | 77.86 | 64.58 | 68.06 | 65.28 | 65.97 | 64.58 | 67.36 |
A05 | 61.43 | 59.03 | 45.14 | 59.72 | 61.11 | 57.64 | 59.03 |
A06 | 71.43 | 65.28 | 56.94 | 65.28 | 65.28 | 65.28 | 65.97 |
A07 | 84.29 | 59.72 | 54.17 | 59.72 | 61.11 | 62.50 | 70.83 |
A08 | 93.57 | 91.67 | 90.97 | 90.28 | 91.67 | 90.97 | 90.97 |
A09 | 80.00 | 85.42 | 87.50 | 85.42 | 86.11 | 85.42 | 90.28 |
Mean | 80.32 | 73.46 | 69.91 | 74.07 | 74.92 | 73.84 | 76.70 |
Std | 10.25 | 15.94 | 18.22 | 15.21 | 15.43 | 15.93 | 15.33 |
\(^{*}p\) value | 0.3047 | 0.2813 | 0.0156 | 0.5313 | 0.0156 |
Subject | Lambda | Session IV | Session V | ||
---|---|---|---|---|---|
Shift-detected (%) | Shift-validated (%) | Shift-detected (%) | Shift-validated (%) | ||
B01 | 0.10 | 10.00 | 4.38 | 6.88 | 3.13 |
B02 | 0.80 | 6.88 | 1.25 | 9.38 | 5.63 |
B03 | 1 | 6.88 | 2.50 | 8.13 | 6.25 |
B04 | 1 | 1.88 | 0.63 | 3.75 | 1.25 |
B05 | 0.30 | 7.50 | 4.38 | 6.88 | 3.75 |
B06 | 0.10 | 8.13 | 5.00 | 10.00 | 6.88 |
B07 | 0.10 | 6.25 | 5.63 | 7.50 | 2.50 |
B08 | 0.20 | 6.25 | 4.38 | 10.00 | 5.00 |
B09 | 0.50 | 8.13 | 4.38 | 8.13 | 3.75 |
4.2 Results for dataset 2B
10-CV | Baseline | SSL |
\(\text {TLCSD}_1 \)
|
\(\text {TLCSD}_2 \)
| ALCSD | UB | |
---|---|---|---|---|---|---|---|
Tr | Eval | Eval | Eval | Eval | Eval | Eval | |
B01 | 70.42 | 69.69 | 66.56 | 69.06 | 70.31 | 71.88 | 75.00 |
B02 | 61.25 | 49.58 | 51.56 | 50.00 | 50.63 | 50.00 | 51.56 |
B03 | 56.67 | 51.56 | 49.38 | 48.44 | 52.81 | 52.81 | 52.19 |
B04 | 88.85 | 93.13 | 85.63 | 93.44 | 93.75 | 93.44 | 96.56 |
B05 | 76.15 | 52.81 | 51.25 | 62.81 | 63.75 | 54.37 | 77.19 |
B06 | 70.71 | 72.81 | 67.50 | 72.19 | 74.06 | 73.13 | 74.06 |
B07 | 84.29 | 58.13 | 56.25 | 59.38 | 61.88 | 62.50 | 70.00 |
B08 | 61.79 | 65.63 | 64.38 | 65.63 | 83.13 | 77.81 | 88.44 |
B09 | 66.25 | 73.75 | 72.19 | 74.38 | 77.19 | 75.00 | 75.00 |
Mean | 70.71 | 65.23 | 62.74 | 66.15 | 69.72 | 67.88 | 73.33 |
Std | 10.78 | 13.98 | 11.89 | 13.64 | 14.05 | 14.16 | 14.67 |
p value | 0.0391 | 0.6719 | 0.0039 | 0.0039 | 0.0039 |