Table 4.
Research articles on pediatric invasive BCIs: signal processing techniques and results (only for pediatric age).
| References | Online/offline/# of classes | Signal processing and features | Classifier/outcome measures | Results |
|---|---|---|---|---|
| Sanchez et al. (2008) | n/a Chance level: n/a |
Fs: 381.5 Hz Filtering: FIR filter (1–6 kHz) Features: (i) Equiripple FIR filter: 1–60 Hz, 60–100 Hz, 100–300 Hz, 300 Hz−6 kHz (ii) FIR filter topology trained using the Wiener solution |
Classifier: n/a Outcome Measure: (i) Pearson's r |
Offline accuracy: n/a Online accuracy: n/a Additional measures: Pearson's r for X-position; Y-position P1: 0.39 ± 0.26; 0.48 ± 0.27 P2: 0.42 ± 0.26; 0.45 ± 0.25 Highest r achieved with 300 Hz−6 kHz feature |
| Breshears et al. (2011) |
Online only
2 classes: Imagined or performed motor movement vs. rest Chance level: 50% (theoretical chance level) |
Fs: 1,200 Hz Filtering: Autoregressive spectral coefficients in 2 Hz frequency bins from 0 to 250 Hz for each electrode Features: (i) Spectral power of filtered frequency bins (ii) Spectral power of electrodes |
Classifier:
Real-time translational algorithm based on the weighted linear summation of the identified features (showing power increases were assigned positive weights, or power decreases were assigned negative weights) Outcome Measure: (i) Accuracy for each action |
Offline accuracy: n/a Online accuracy: P1: 70.8–99.0% P2: 72.7–77.4% P3: 82.7–85.1% P4: 75.0–100% P5: 88.8 % P6: 93.3 % Additional measures: n/a |
| Pistohl et al. (2012) |
Offline
2 classes: precision grip, whole-hand grip 10-fold cross-validation (20 repetitions) Chance level: 50% (theoretical chance level) |
Fs: 256 Hz Filtering: Re-referenced to common average, average voltage subtracted, normalized voltage, low pass filtered component (fc ~5 Hz) Features: (i) Signal components |
Classifier: rLDA Outcome Measure: (i) Decoding accuracy (ii) Temporal evolution of grasp discriminability |
Offline accuracy:
P1: 97% P2: 84% P3: 96% Online accuracy: n/a Additional measures: Temporal evolution of grasp discriminability: 0.2 s |
| Pistohl et al. (2013) |
Offline only
2 classes: grasp and no grasp Chance level: n/a |
Fs: 256 or 1,024 HzFiltering:LFC: 2nd order Savitzky-Golay filter (window length: 250 ms) Features: (i) LFC (ii) Frequency band amplitudes within consecutive bands of 4 Hz width from 0 to 128 Hz. Band pass filtering by 4th order elliptic digital filter design |
Classifier: rLDA in 10-fold cross-validation Outcome Measures: (i) True positive ratio (TPR) (ii) False positive ratio (FPR) (iii) False positive min−1 (FP-rate) |
Offline accuracy:
After 0.25 s (TPR/FPR/FP-rate): P1: 0.75/0.26/2.5 P2: 0.50/0.36/2.7 P3: 0.75/0.25/3.1 After 0.50 s (TPR/FPR/FP-rate): P1: 0.92/0.10/0.9 P2: 0.69/0.12/0.9 P3: 0.91/0.08/1.0 After 0.75 s (TPR/FPR/FP-rate): P1: 0.97/0.05/0.4 P2: 0.74/0.05/0.4 P3: 0.96/0.03/0.4 Online accuracy: n/a Additional measures: n/a |
Fs, sampling frequency; fc, cut-off frequency; rLDA, regularized linear discriminant analysis; FPR, false positive ratio; TPR, true positive ratio; FP-rate, false positive rate; P#, pediatric participant number; LFC, low-pass filtered component; FIR, finite impulse response; n/a, not applicable.