Beyond rates: time-varying dynamics of high frequency oscillations as a biomarker of the seizure onset zone

Abstract

Objective. High frequency oscillations (HFOs) recorded by intracranial electrodes have generated excitement for their potential to help localize epileptic tissue for surgical resection. However, the number of HFOs per minute (i.e. the HFO 'rate') is not stable over the duration of intracranial recordings; for example, the rate of HFOs increases during periods of slow-wave sleep. Moreover, HFOs that are predictive of epileptic tissue may occur in oscillatory patterns due to phase coupling with lower frequencies. Therefore, we sought to further characterize between-seizure (i.e. 'interictal') HFO dynamics both within and outside the seizure onset zone (SOZ).Approach. Using long-term intracranial EEG (mean duration 10.3 h) from 16 patients, we automatically detected HFOs using a new algorithm. We then fit a hierarchical negative binomial model to the HFO counts. To account for differences in HFO dynamics and rates between sleep and wakefulness, we also fit a mixture model to the same data that included the ability to switch between two discrete brain states that were automatically determined during the fitting process. The ability to predict the SOZ by model parameters describing HFO dynamics (i.e. clumping coefficients and coefficients of variation) was assessed using receiver operating characteristic curves.Main results. Parameters that described HFO dynamics were predictive of SOZ. In fact, these parameters were found to be more consistently predictive than HFO rate. Using concurrent scalp EEG in two patients, we show that the model-found brain states corresponded to (1) non-REM sleep and (2) awake and rapid eye movement sleep. However the brain state most likely corresponding to slow-wave sleep in the second model improved SOZ prediction compared to the first model for only some patients.Significance. This work suggests that delineation of SOZ with interictal data can be improved by the inclusion of time-varying HFO dynamics.

Keywords: epilepsy; epileptogenic zone; hierarchical Bayesian methods; high-frequency oscillations (HFOs); intracranial EEG; ripple; surgery.

Figure 1:

Modulation of HFO rate over 24 hours in one example patient. HFOs were automatically detected using the algorithm by Charupanit and Lopour (2017). Automated HFO counts per minute are shown from an iEEG channel in the left parahippocampal cortex (blue line) and an iEEG channel in the medial temporal gyrus (green line), with HFO counts per minute denoted by the right y-axis. These counts are overlayed on a grey-scale color map of low frequency band power (1–20 Hz) from the same left parahippocampal iEEG channel during the same 24 hour time period. Darker shading indicates higher power in the frequency band denoted by the left y-axis. HFO counts (blue and green lines) are modulated by the sleep-wake cycle (von Ellenrieder et al., 2017), as evidenced by their correlation with delta (1–4 Hz) frequency power (grey-scale color map). HFOs are typically analyzed during slow-wave sleep. However, with the method presented here, the data does not need to be visually sleep staged prior to classification of SOZ and non-SOZ channels.

Figure 2:

Top: Simulated instantaneous HFO rates (per second) from a Negative Binomial process with a small clumping coefficient (ζ = 0.01, near a Poisson process) over a 60 minute period. Electrodes that contained Negative Binomial processes with small clumping coefficients were found to be predictive of SOZ. Middle: Simulated instantaneous HFO rates from a Negative Binomial Process with some clumping (ζ = 1). Bottom: Simulated instantaneous HFO rates from a Negative Binomial Process with significant clumping (ζ = 10). All simulations had a HFO rate of λ = .1 per second.

Figure 3:

SOZ prediction based on the clumping coefficients (CC) estimated by Model 1 (Top) and Model 2 (Bottom). (Left) ROC curves when using small values of the CC to identify SOZ channels. ROC curves for individual patients (N = 16) are displayed using fine lines, and the average is shown in bold. The bold dashed line indicates an ROC at chance prediction. Data points on the top of these two plots indicate the false positive rate (FPR) for which the true positive rate (TPR) is 1, with patient labels to compare across plots. From Model 2, the brain state with the most delta (1–4 Hz) power in each patient was labeled State A (green lines), while the other model-found brain state was labeled State B (blue lines). (Right) Distribution of AUC values based on CC for each patient. The exact AUC values are denoted as hexagons or stars on the x-axis, while the shaded distributions are a density approximation from N = 16 values. Hexagons denote patients where the analysis was performed exclusively on grey matter channels, while stars denote patients for which all channels were included. Each point is labeled with the corresponding patient number, and the y-values are sorted by AUC; however, the y-values have no other meaning. The bold dashed lines indicate an AUC of 0.5.

Figure 4:

The aggregate ROC curves for all included channels (total channel count of 1391) across all patients (N = 16) when using small values of the HFO clumping coefficient (CC) estimates to identify SOZ channels. A few representative cutoff CC values are shown in the text boxes, such that values smaller than the CCs in text boxes generated the indicated points on the ROC curves across all included patients. (Left) Aggregate ROC curve generated from CC estimates using Model 1 (CC1). (Right) Aggregate ROC curve generated from CC estimates using Model 2. The brain state with the most delta (1–4 Hz) power in each patient was labeled brain state A while the other model-found brain state was labeled brain state B. The green curve was generated from all CCs for each channel from all patients’ model-found brain states A (CC2A). The blue curve was generated from all CCs for each channel from all patients’ model-found brain states B (CC2B). The bold dashed lines indicate an AUC of 0.5.

Figure 5:

SOZ prediction based on the coefficients of variation (CV) estimated by Model 1 (Top) and Model 2 (Bottom). (Left) ROC curves when using small values of the CV to identify SOZ channels. (Right) Distribution of AUC values based on CV for each patient. Readers are referred to the caption of Figure 3 for a detailed description of the plot elements.

Figure 6:

SOZ prediction based on the HFO rate estimated by Model 1 (Top) and Model 2 (Bottom). (Left) ROC curves when using large values of the HFO rate to identify SOZ channels. (Right) Distribution of AUC values based on HFO rate for each patient. Readers are referred to the caption of Figure 3 for a detailed description of the plot elements.

Figure 7:

Model-derived brain states correspond to sleep and wakefulness. Representative examples are shown from Patients 1, 11, 15, and 16 with brain states A (dark green dots) and B (dark blue dots) obtained automatically every 5 minutes from Model 2. The labels of states A and B were assigned using the mean slow-wave delta power (1–4 Hz; standardized mean across channels), with brain state A containing higher delta power. Black lines represent the standardized mean delta power across channels. In patients 15 and 16, who had concurrent iEEG and scalp EEG, the HFO model-derived brain states differentiated slow wave sleep (i.e. NREM sleep stages 1, 2 and 3; denoted by light green dots in the upper portion of the bottom two subplots) from all other states (REM sleep and wakefulness; denoted by light blue dots in the lower portion of the bottom two subplots). This determination was made based on a comparison to expert sleep staging.