A Novel Time-Incremental End-to-End Shared Neural Network with Attention-Based Feature Fusion for Multiclass Motor Imagery Recognition

Figure 10

Confusion matrix of the trained network model. The classification accuracy rate of each subject in BCI Competition IV 2a (A01–A09). The average classification accuracy rate of all subjects (average). Take the confusion matrix of A01 as an example. The first column indicates that, in a total of 72 testing samples from 72 left-hand trials, the number of testing samples is 64 whose model predicted value of classification category is left hand, and the number of testing samples is 8 whose model predicted value of classification category covers right hand, feet, and tongue; that is, TP = 64 and FN = 8; then, TPR = 64/72 = 88.9%. The second to fourth columns are the same. The first row indicates that, in a total of 72 testing samples from 64 left hand, 2 right hand, 5 feet, and 1 tongue trials, all the testing samples’ model predicted values of classification category are left hand; that is, TP = 64 and FP = 8; then, PPV = 64/72 = 88.9%. The second to fourth rows are the same. The main diagonal indicates that the total number of model’s correct judgments is 64 + 62+64 + 64 = 254 times, while the total model prediction results of testing samples are 288, so, ACC = 254/288 = 88.2%.