Machine-learning Assisted Differentiation Between Physiological Oscillation Epochs and Muscle or Epileptiform activity in EEGs
Vikram Jakkamsetti1, Gauri Kathote1, Adrian Avila1, Sharon Primeaux1, Aksharkumar Dobariya1, Ignacio Malaga-Dieguez1, Juan Pascual1
1University of Texas Southwestern
Objective:
Machine-learning assisted separation of spontaneous cognition-relevant EEG oscillations from EEG perturbation induced by epileptiform events and muscle artifacts.
Background:
Transient band-limited alpha and gamma frequency oscillations in EEGs of awake subjects reflect cognitive processes and integrity of excitation-inhibition balance. These oscillations are diminished in disorders such as mitochondrial encephalopathies as we have previously shown. Automated capture of such oscillation epochs in prolonged EEGs can prove valuable for numerous clinical and research studies. However, EEG recordings can be degraded by superimposed high frequency activity inherent in movement and epileptiform discharges.
Design/Methods:
Forty band-limited alpha and gamma spontaneous oscillation epochs were recorded from electrode C3 in a standard 20-lead EEG configuration. The gaussian distribution of non-event wavelet scalogram allowed automated identification of outlier putative events, which were then confirmed or rejected by a human observer. An instantaneous wavelet scalogram “cross-section” of the event was used to generate a representative frequency waveform for that event and analyzed using action potential waveform sorting Wave clus (Neural Computation 16, 1661-1687; 2004). This protocol was applied to EEGs containing muscle (27 events) and to EEGs with epileptiform activity (36 events) due to mitochondrial pyruvate dehydrogenase deficiency.
Results:
Oscillation epoch waveforms exhibited distinct band-limited peaks which were absent for muscle or epileptiform activity. Muscle events uniformly displayed broad-band high-frequencies, whereas epileptiform events were more variable. Oscillation epoch wavelet coefficients were reliably differentiated from other events (MANOVA p<10-4). A support vector machine-learning protocol examining the difference between epochs and muscle and epileptiform events demonstrated a misclassification rate of zero and 1.3%, respectively, and an area under the curve for a receiver operating characteristic curve of 1 for both.
Conclusions:
Instantaneous wavelet scalogram profiles provide high precision in differentiating oscillation epochs from muscle or epileptiform activity, allowing for automated oscillation epoch detection in subjects with epilepsy.