Skip to content

Features (Proposed)

scottclowe edited this page Nov 10, 2014 · 1 revision

Here are two papers describing some useful methods Pereda, 2005 and Wulsin et al. 2011

I've parsed some of the discussion of the previous challenge (differentiating interictal and ictal iEEGs on almost identical datasets). While this was an easier challenge (seizure or not seizure is clearer than pre-seizure or not seizure) it still offers a lot of ideas. It seems all the best performing entries used subject-specific models.

Ideas of things to do and features to look for:

  • Variance measures:
    • variance
    • max-min
    • 95th-5th percentile
    • standard deviation
    • For each channel, over channels and for cross-correlations)
  • Covariance or Correlation coefficient
    • and its eigenvalues
  • Power spectral density:
    • Logarithmically downsampled raw PSD curve
    • Average spectral power (wavelets) at a range of bands Delta (0/0.5-4Hz), Theta (4-7/8Hz), Alpha (7/8-14Hz), Beta (14/15-30Hz), Low Gamma (30-100Hz) High Gamma (100-200Hz)
    • Ratio of Spectral Powers to each other (e.g., Theta / Alpha; all combinations).
  • Coherence spectral density
  • Cross-correlation
    • peak value & lag
    • width
  • Partial autocorrelation function
  • Cross-frequency coupling
  • Nonlinear correlation coefficient (Zaveri et al., 1992)
  • Phase-synchrony (Chavez et al., 2003)
  • Auto-regressive modelling
    • Granger Causality
  • Transfer Entropy
  • Independent Component Analysis
  • Empirical mode decomposition
  • Fractal dimension
  • Mobility and complexity measures
  • Skewness and kurtosis
  • log sum of absolute values for each channel
  • log of first 250 values of FFT for channels [isn't log(abs(fft(data))) basically the same as psd(data)?]
  • Correlation coefficient for FFT
  • Models based on Wulsin et al. 2011 (Appendix B?) but note didn't perform better than simpler features in previous competition but that was using fairly simple Scikit ExtraTreesClassifier
    • Normalized positive area under curve
    • Normalized decay (chance-corrected fraction of data that is increasing or decreasing)
    • Line length (sum of abs difference between successive samples)
    • Mean energy: mean of the square of samples across each channel
    • Average peak amplitude: base-10 log of the mean-squared amplitude of each peak in the channel
    • Average valley amplitude: as above but for the valleys
    • Normalized peak number: number of peaks in the channel normalized by the mean difference between adjacent data points
    • Zero crossings: subtracts the mean value (rather than the line of best fit as in the paper) from the channel and then counts how many times the zero-mean data crosses zero

Winner of the previous challenge posted his methodologies and code. They did "FFT 1-47Hz, concatenated with correlation coefficients (and their eigenvalues) of both the FFT output data, as well as the input time data. The data was then trained on per-patient Random Forest classifiers (3000 trees)"

2nd (I've presumed) and 3rd also posted their code.

All of the above solutions need further investigation

The amusing solution previously side-stepped feature selection, converted signals into images then used 'nolearn' image classification. Entertaining alternative explained by them in more detail here

Pereda, E., Quiroga, R. Q., & Bhattacharya, J. (2005). Nonlinear multivariate analysis of neurophysiological signals. Progress in neurobiology, 77(1), 1-37.

Chavez, M., Le Van Quyen, M., Navarro, V., Baulac, M., Martinerie, J., 2003. Spatio-temporal dynamics prior to neocortical seizures: amplitude versus phase couplings. IEEE Trans. Biomed. Eng. 50, 571–583

Zaveri, H. P., Williams, W. J., Iasemidis, L. D., & Sackellares, J. C. (1992). Time-frequency representation of electrocorticograms in temporal lobe epilepsy. Biomedical Engineering, IEEE Transactions on, 39(5), 502-509.

D F Wulsin, J R Gupta, R Mani, J A Blanco and B Litt, Modeling electroencephalography waveforms with semi-supervised deep belief nets: fast classification and anomaly measurement, J. Neural Eng. 8 (2011) 036015

Clone this wiki locally