1 Introduction
2 Materials and methods
2.1 Dataset
2.2 Feature extraction
EEG features |
---|
Total power (0–12 Hz) |
Peak frequency of spectrum |
Spectral edge frequency (SEF 80 %, SEF 90 %, SEF 95 %) |
Power in 2 Hz width subbands (0–2, 1–3,…10–12 Hz) |
Normalized power in same subbands |
Wavelet energy (Db4 wavelet coefficient corresponding to 1–2 Hz) |
Curve length |
Number of maxima and minima |
Root mean square amplitude |
Hjorth parameters (activity, mobility and complexity) |
Zero crossing rate (ZCR), ZCR of the Δ and the ΔΔ |
Variance of Δ and ΔΔ |
Autoregressive modelling error (AR model order 1–9) |
Skewness and kurtosis |
Nonlinear energy |
Shannon entropy—spectral entropy, singular value decomposition entropy |
Fisher information |
Linear filter bank: 15 subband energies (0–2, 1–3,…14–16 Hz) |
15 Cepstral coefficients |
15-s order frequency filtered bank energies |
Peak–peak voltage |
2.3 Feature baseline correction (FBC)
2.4 Training datasets and test procedure
2.5 SVM classifier training
A: Adults test set | No FBC | FBC | Performance increase (%) |
C, σ
|
---|---|---|---|---|
ACP |
0.85 ± 0.005
|
0.93 ± 0.002
| 52 | 40, 7 |
NCP
|
0.84 ± 0.014
|
0.92 ± 0.005
| 52 | 40, 5 |
CCP
|
0.83 ± 0.04
|
0.93 ± 0.009
| 57 | 50, 6 |
B: Neonatal test set | No FBC | FBC | Performance increase (%) |
C, σ
|
---|---|---|---|---|
ACP | 0.70 ± 0.004*^ | 0.86 ± 0.002*^ | 54 | 20, 8 |
NCP
|
0.76 ± 0.042*
|
0.90 ± 0.009*
| 59 | 20, 8 |
CCP |
0.78 ± 0.028^
|
0.90 ± 0.009^ | 53 | 50, 6 |