Extended Data Fig. 4: Overall performance of the ErrP detector.
From: A memristor-based adaptive neuromorphic decoder for brain–computer interfaces
a, Average TPR, TNR and BACC as a function of the number of samples (n = 100, 200, 300, 500, 700, and 1000) in the training set, and the testing set contains 360 samples. The samples in the training set and testing set are randomly selected from the data collected in the ErrP calibration experiment without overlap. b, Average TPR, TNR and BACC as a function of the classification threshold. If the feature value of the DCPM-based classifier exceeds a predefined threshold, it labels the classification of “target”; otherwise, it labels the classification of “non-target”. These metrics are calculated from the leave-one-block-out cross-validation on the training data. c, Confusion matrix between the “target” class and the “non-target” class when the threshold is set as 0.5 in b. d, Grand averaged metrics on 10 blocks online data from all 10 subjects (n = 100). The error bars in a and d, and the shaded areas in b indicate the mean values ± s.d. e, Correct and incorrect evoked signals from channel FCZ (green and red lines, respectively). The shade areas in e indicate the 95% confidence intervals.
