Noninvasive classification of physiological and pathological high frequency oscillations in children

Figure 1. HFO characterization through temporal/morphological, spectral and spatial features. (A) Each panel (from left to right) show: (i) EEG signals filtered between 80 and 160 Hz on an expanded timescale of 400 ms (ripple time window in yellow); (ii) time–frequency analysis showing the ripple ‘island’ (isolated spectral peak) based on the power ratio computation; and (iii) source localization, on a subject-specific cortical surface, of the HFO, showing the active real sensors (channels) and cortical sources (vertices). (B) Comprehensive list of features and their descriptions. Temporal/morphological (green), spectral (blue) and spatial (red). (C) Classification of HFOs from DRE patients. The asterisk represents the maximum value of the wMEM map located in the epileptogenic (red) or non-epileptogenic (orange) region. Ch, channel; EpiR, epileptogenic region; Freq, frequency; HFO, high frequency oscillation; L, left; non-EpiR, non-epileptogenic region.

Figure 2. Flow diagram of participants’ inclusion in the study. The diagram shows the number of participants recruited for this study for both the DRE and TDC group, the number of sleep recordings acquired with each modality, the number of participants where HFOs were detected and their respective epilepsy classification. DRE, drug-resistant epilepsy; EEG, electroencephalography; HD, high-density; HFO, high frequency oscillation; MEG, magnetoencephalography; TDC, typically developing controls.

Figure 3. Examples of HFOs recorded with HD-EEG and MEG. (A) HFOs recorded from two TDC (6- and 12-year-old females). (B) HFOs recorded from two children with DRE (Patients 47 and 49 in Supplementary Table 1). Each panel shows one HFO recorded with HD-EEG (top) and one with MEG (bottom). Within each panel, we report (from left to right): (i) the signals filtered between 80 and 160 Hz on an expanded timescale (ripple time window in red); (ii) time–frequency analysis showing ripple ‘island’ (isolated spectral peak); and (iii) source localization, on the subject's 3D cortical surface, with maximum activation outlined in yellow. DRE, drug-resistant epilepsy; EEG, electroencephalography; HD, high-density; HFO, high frequency oscillation; L, left; MEG, magnetoencephalography; R, right; TDC, typically developing controls.

Figure 4. HFOs on HD-EEG (orange) and MEG (azul) recorded from the TDC (blue) and DRE group (red). (A) Proportions of subjects presenting HFOs in the different groups are displayed via pie charts. (B) Comparisons between HFO rates in TDC and DRE children are displayed via boxplots. (C) Correlation of HFO rate with the subjects’ age (Pearson correlation). Regression lines are shown for DRE (red). r = correlation coefficient. *P < 0.05; **P < 0.01; ***P < 0.001. DRE, drug-resistant epilepsy; EEG, electroencephalography; HD, high-density; HFO, high frequency oscillation; MEG, magnetoencephalography; TDC, typically developing controls.

Figure 5. Physiological HFOs from TDC: HD-EEG findings. (A) Boxplots of HFO rates (HFOs/min) by cortical region and associated distribution projected onto a generic cortical surface colour-coded based on the mean rate per lobe (HFOs/min). (B) HFO feature comparisons between regions. Only features showing significant different are shown. Cingulate and occipital regions were excluded due to low HFO rates (median equal to zero). (C) Correlation of HFO features with participants’ age. Each data point represents the feature mean between all HFOs recorded for a patient. r = correlation coefficient. *P < 0.05; **P < 0.01; ***P < 0.001. EEG, electroencephalography; HD, high-density; HFO, high frequency oscillation.

Figure 6. HD-EEG HFO feature comparison. (A–H) Comparison of each feature between four classes of HFOs: HFOs coming from the healthy brain (blue) of TDC, epileptogenic (EpiR, red) and non-epileptogenic (non-EpiR, orange) regions of DRE patients and patients with diffuse or generalized DRE (diffuse DRE, grey). Dash-line boxplots show results obtained using only DRE patients that had good surgical outcome after surgery (where non-EpiRs are unambiguously defined). DRE, drug-resistant epilepsy; EEG, electroencephalography; EpiR, epileptogenic region; HD, high-density; HFO, high frequency oscillation; non-EpiR, non-epileptogenic region; TDC, typically developing controls.

Figure 7. Naïve-Bayes classification of HFOs recorded with HD-EEG (10-fold cross-validation). (A) ROC curves (one-versus-all binary classification) for each class of HFO (blue: TDC HFO versus others; orange: non-EpiR HFO from DRE versus others; red: EpiR HFO from DRE versus others). (B) Confusion matrix obtained for the classification of the HFOs recorded from the DRE group (N = 266) into epileptogenic versus non-epileptogenic. ACC, accuracy; AUC, area under the curve; DRE, drug-resistant epilepsy; EpiR, epileptogenic region; HFO, high frequency oscillation; non-EpiR, non-epileptogenic region; NPV, negative predictive value; PPV, positive predictive value; ROC, receiver operating characteristics; TDC, typically developing controls.