Measuring Real-Time Medication Effects From Electroencephalography

Supplemental Digital Content is Available in the Text.

Abstract

Purpose:

Evaluating the effects of antiseizure medication (ASM) on patients with epilepsy remains a slow and challenging process. Quantifiable noninvasive markers that are measurable in real-time and provide objective and useful information could guide clinical decision-making. We examined whether the effect of ASM on patients with epilepsy can be quantitatively measured in real-time from EEGs.

Methods:

This retrospective analysis was conducted on 67 patients in the long-term monitoring unit at Boston Children's Hospital. Two 30-second EEG segments were selected from each patient premedication and postmedication weaning for analysis. Nonlinear measures including entropy and recurrence quantitative analysis values were computed for each segment and compared before and after medication weaning.

Results:

Our study found that ASM effects on the brain were measurable by nonlinear recurrence quantitative analysis on EEGs. Highly significant differences (P < 1e-11) were found in several nonlinear measures within the seizure zone in response to antiseizure medication. Moreover, the size of the medication effect correlated with a patient's seizure frequency, seizure localization, number of medications, and reported seizure frequency reduction on medication.

Conclusions:

Our findings show the promise of digital biomarkers to measure medication effects and epileptogenicity.

Antiseizure medications (ASMs) are the mainstay of epilepsy treatment, aimed at achieving and maintaining seizure freedom. Many factors inform ASM selection, including patient's age, sex, epilepsy syndrome, seizure types and burden, presence of structural brain abnormalities, and comorbidities, together with the drug properties, side effects, and possible drug–drug interactions.^1–7 Once a medication regimen has been started, the primary measure of efficacy is ongoing seizures, or resolution, which is clinically assessed by diaries, often difficult to ascertain, and may take weeks or even years of monitoring.^8–10 ASM monotherapies and polytherapies are often trialed in an attempt to achieve seizure control.^3,4,6 With adequate pharmacological control of seizures, patient outcomes can improve and epilepsy-associated morbidity and mortality may be reduced.¹¹ However, one-third of patients with epilepsy (PWE) do not respond well enough or are resistant to the medications.¹² Evaluation of the initial response to medication may help to predict long-term outcomes in newly diagnosed patients.^13,14 This may also be useful for early medication decisions because there are also concerns regarding the deleterious effects of some ASMs on the developing brain.¹⁵

The current approach to epilepsy management has many limitations. (1) Evaluating response to medication is slow.¹⁶ (2) Seizure diaries can often be unreliable.¹⁷ (3) There are no independent, objective, and quantitative biomarkers to guide epilepsy management and treatment. The difficulty in choosing a medication is exacerbated by the complexity of individual patient differences in drug response.

An objective biomarker for ASM effects will be a valuable tool for the treatment of epilepsy because it provides a noninvasive, real-time, and quantitative measurement of medication effects on epileptogenicity and can guide clinical decision-making. As an early step in this direction, we evaluate changes in epilepsy patients' brain neurophysiology while the medication regimen is changed during inpatient presurgical evaluation. This study uses multifrequency nonlinear measures computed from EEG recordings to characterize functional brain dynamics in epilepsy. Our central hypothesis is that lower complexity in EEG signals will be associated with increased epileptogenicity or tendency to have seizures because ASMs are reduced. This will be indicated by changes in nonlinear dynamical EEG measures. These measures may be useful as digital biomarkers to provide information regarding ASM effect on the brain and treatment response in PWE. The goal of this study was to take first steps toward potential digital biomarkers that can be used as decision support tools to provide clinicians with immediate insight into whether a medication is useful for a particular patient from the analysis of a relatively short EEG segment. Paired with the targeted therapy, a digital biomarker may provide quick and quantitative measures of medication response. Although considerably more research is required for a validated clinical biomarker, the results of this study show a promising direction for this endeavor.

METHODS

Study Population

This study was approved by the Institutional Review Board (IRB) at Boston Children's Hospital (BCH) (IRB-P00001945). The study design was a retrospective deidentified record review; therefore, the need for informed consent has been waived the BCH IRB. The research was performed in accordance with the guidelines and regulations of the IRB at BCH and all applicable government regulations, including the Helsinki guidelines.

We designed our study around PWE admitted for long-term EEG monitoring who underwent medication weaning. A retrospective chart review of patients admitted to the long-term monitoring (LTM) unit at BCH between July 2016 and June 2017 was performed. We included all patients admitted for presurgical evaluation and on a medication regime of at least one ASM before admission. Patients who did not have an epilepsy diagnosis, patients who had seizure clusters (four or more seizures in 15 minutes), and patients whose EEG or clinical data were not retrievable were excluded. To evaluate seizure burden across patients, we considered seizure frequency (daily, weekly, monthly, and annually) and the number of seizures that occurred in the past 30 days before admission as reported by caregivers. Response to medication was defined as any clinically determined improvement in seizure burden, namely the reported reduction in seizure frequency, after medication initiation.

EEGs were measured with a Natus NeuroWorks XLTEK EMU 40 system and reviewed with NeuroWorks software version 8, using a standard 10 to 20 system and common references A1 and A2. The sampling rate was 256 to 1,024 Hz; all signals were subsampled to a common sampling rate of 256. No other filtering was performed. Leads AF3 and AF4 were analyzed as replacement channels for Fp1 and Fp2 in some patients at risk of skin lesions in the setting of electrode placement. An epileptologist reviewed the EEGs and reports to obtain electrographic seizure onset, offset, and seizure type. Seizure start and end times were determined for clinical care and then rereviewed by a clinical fellow for this study.

ASMs included common antiepileptic drugs and benzodiazepines. In addition to ASMs, CBD, and benzodiazepines, we similarly accounted for other medications that may potentially interfere with an EEG activity, including pyridoxine, prednisone, acetazolamide, propofol, lecithin, hydrocortisone, magnesium, theanine, zinc, and somatropin^18–20 (see Table S1, Supplemental Digital Content 1, http://links.lww.com/JCNP/A193). The ASM-high periods were defined as the first 30-second seizure-free interval of recorded EEG available 30 minutes from the start of the EEG recording. These segments were chosen to characterize patients' standard epileptogenic potential while on their baseline level of ASM during admission. The ASM-low period was defined as the first 30-second seizure-free interval of recorded EEG from the end of the largest gap in ASM administration during the patients' admission. The ASM-low periods were chosen to characterize the patients' epileptogenicity when ASMs are tapered or weaned. Because we were identifying the largest gap in all medication being administered, it did not matter whether the dosage was decreased or whether a medication was dropped. No medication was administered during that time. No seizures occurred one hour before or one hour after the selected EEG intervals. We did not account for state of vigilance when choosing the segments, although all EEG segments were from awake patients.

We selected two 30-second EEG samples from each patient. EEG samples were indicative of an ASM-high and ASM-low period. Thirty-second EEG segments from these patients were retrospectively selected for analysis. This segment length has been demonstrated to be appropriate in our previous studies of EEG biomarkers for epilepsy and autism.^21–23 A preliminary study has demonstrated that continuous 20-second segments of EEG recording are sufficient for computing nonlinear measures,²⁴ while other studies have also found that EEG segments of 60 seconds or less are stable for nonlinear analysis.²⁵ Recurrence quantification analysis (RQA), a method of nonlinear time series analysis for the investigation of dynamical systems,^26,27 was used to compute dynamical properties from each EEG signal. Power was also computed. Recurrence plots are graphical representations of the phase portrait of a dynamical system, and the quantitative values derived by RQA, in principle, contain information about all the essential dynamics of the system, including measures that are equivalent to analytical formulations of entropy, Lyapunov exponents, and other measures.^26,27 More detailed discussions of RQA measures in the context of epilepsy can be found in previous publications.^21,23,28 Nonlinear measures derived from recurrence plots and used in this article include recurrence rate (RR), determinism (DET), trapping time (TT), line length maximum (Lmax), mean (Lmean), and entropy (Lentr).

The EEG signal from each of the 19 electrodes on the standard 10 to 20 EEG system was decomposed into seven power-of-two frequency bands using the DB4 wavelet transform. Wavelet details containing discrete frequency bands were reconstructed to yield discrete frequency subsignals²² that correspond approximately to commonly used frequency bands: delta (0–4 Hz), theta (4–8 Hz), alpha (8–16 Hz), beta (16–32 Hz), gamma (32–64 Hz), and gamma+ (64–128 Hz). RQA^29,30 was used to compute nonlinear measures from each EEG sensor, following the procedures described in previous publications.^21,23 These metrics characterize different aspects of the essential dynamics of a complex system.^23,31,32 Power was also computed on each band.

All statistical analysis was performed using the R statistical computing package. Wilcoxon rank-sum tests were used to compare group differences across nonlinear features shown in the figures and accompanying tables.

RESULTS

After retrospective review, 67 of 831 patients admitted to the long-term monitoring unit at Boston Children's Hospital between July 2016 and June 2017 met our inclusion criteria (Fig. 1). We included all patients admitted for presurgical evaluation and on a medication regime of at least one ASM started before admission. Patients who did not have an epilepsy diagnosis, patients who had seizure clusters (four or more seizures in 15 minutes), and patients whose EEG or clinical data were not retrievable were excluded. Our cohort (median age 10.2 years, 49.3% female patients) included patients with a variety of epilepsy syndromes, etiologies, and MRI findings. Table 1 presents the clinical characteristics and demographic data of the patients.

FIG. 1.

Inclusion/exclusion flowchart. A flowchart for inclusion and exclusion of patients in this study is shown. 67 of 831 patients admitted to the long-term monitoring unit at Boston Children's Hospital between July 2016 and June 2017 met our inclusion criteria.

TABLE 1.

Clinical Characteristics of the Study Population (n = 67)

Demographic Characteristics	Median Years, (Range, IQR)
Female, n (%)	33 (49.3)
Median age at first seizure	2.3 (11.6, 5.2)
Median age at EEG	10.2 (22.5, 6.5)
Ethnicity, n (%)
Not Hispanic or Latino	47 (70.1)
Not reported	9 (13.4)
Unknown	6 (9.0)
Hispanic or Latino	5 (7.5)
Race, n (%)
White	41 (61.2)
Not reported	14 (20.9)
Unknown	8 (11.9)
Black or African American	3 (4.5)
Asian	1 (1.5)
Clinical characteristics
Epilepsy syndrome, n (%)
None	48 (71.6)
Temporal lobe epilepsies	6 (9.0)
Frontal lobe epilepsies	4 (6.0)
West syndrome (infantile spasms)	1 (1.5)
Sturge Weber syndrome	1 (1.5)
Neurofibromatosis	1 (1.5)
Tuberous sclerosis complex	1 (1.5)
Electrical status epilepticus in sleep (ESES)	1 (1.5)
Continuous spikes and Waves during sleep (CSWS)	1 (1.5)
Familial focal epilepsy with variable foci (FFEVF)	1 (1.5)
Dravet syndrome	1 (1.5)
Other undetermined epilepsy	1 (1.5)
Comorbidities*, n (%)
Developmental/psychiatric disorders	37 (55.2)
Other medical conditions	26 (38.8)
None	19 (28.4)
Neurological disorders	18 (26.9)
Manifestation of a seizure causing multisystem genetic disorder	2 (3.0)
Surgical history, n (%)
Neurosurgery	8 (11.9)
MRI conducted 6 months prior/after EEG of interest, n (%)
Yes	66 (98.5)
No	1 (1.5)
MRI findings*, n (%)
Dysplasia	16 (23.9)
Encephalomalacia/trauma	13 (19.4)
Gliosis, unspecified	12 (17.9)
Normal	11 (16.4)
Hippocampal sclerosis	10 (14.9)
Volume loss, unspecified	10 (14.9)
Nonspecific finding (not seizure related)	10 (14.9)
Tumor	6 (9.0)
Resection	5 (7.5)
Cyst	4 (6.0)
Degeneration	3 (4.5)
Malformation	3 (4.5)
Infarction	3 (4.5)
Tuberous sclerosis/hamartoma	2 (3.0)
Information unavailable	2 (3.0)
Postsurgical changes	2 (3.0)
Prominence (unspecified significance)	2 (3.0)
A-V malformation/cavernoma/angioma	1 (1.5)
Sturge Weber	1 (1.5)
Encephalocele	1 (1.5)
Hypomyelination	1 (1.5)
Microcephaly	1 (1.5)
Ventriculomegaly	1 (1.5)
Aqueductal stenosis	1 (1.5)
Hypogenesis	1 (1.5)
MRI lesion location*, n (%)
Frontal lobe	12 (17.9)
Temporal lobe (mesial temporal)	11 (16.4)
Temporal lobe (neocortical or unspecified temporal)	9 (13.4)
Hemisphere	9 (13.4)
Parietal lobe	7 (10.4)
Hippocampus	6 (9.0)
Cerebellar	4 (6.0)
Insula	3 (4.5)
Periventricular	3 (4.5)
Cerebral	3 (4.5)
Lateral ventricles	2 (3.0)
Occipital lobe	2 (3.0)
Corpus callosum	1 (1.5)
Multilobar	1 (1.5)
Perisylvian	1 (1.5)
Unknown	1 (1.5)
MRI abnormality lateralization, n (%)
Right	21 (39.6)
Left	20 (37.7)
Bilateral	12 (22.6)
CT scan conducted 6 months prior/after EEG of interest, n (%)
No	66 (98.5)
Yes	1 (1.5)
CT scan abnormalities, n (%)
Yes	1 (100)
CT scan findings, n (%)
Other	1 (100)

^*Patients may have been represented in more than one category, and numbers, therefore, do not add up to 67 (100%). Values are n (%) unless otherwise indicated.

ASM, antiseizure medication; CT, computed tomography.

Effects of Antiseizure Medication on the Brain

We observed a group-level difference across all RQA measures between patients' premedication and postmedication weaning: Wilcoxon test results comparing drug on and drug off groups, all frequencies, and all sensor locations resulted in highly significant P values even after correction for TT (P < 2.2e-16), Lmean (P < 2.2e-16), Lmax (P < 3.3e-14), and Lentr (P < 2.0e-13). Figure 2 illustrates the multifrequency curves for each nonlinear measure across all channels for all patients. The separation between the two multifrequency curves indicates the effect that medication has on the brain across all frequency bands, as measured by the particular nonlinear measure. Figure 3 shows the size of the difference for each measure by computing the percentage change in the areas under the multifrequency curve. TT showed the largest medication effect, followed by the diagonal line length values (Lmean, Lmax, and Lentr), determinism, and RR. The separation between the curves is frequency-dependent. Power was also computed for each of the subsignals and was not found to be significantly different.

FIG. 2.

Multifrequency graphs for patients premedication and postmedication weaning. The multifrequency graphs of recurrence rate (RR), determinism (DET), trapping time (TT), line length maximum (Lmax), mean (Lmean), and entropy (Lentr). The solid red curve represents the electrodynamics of the brain for patients on medication, and the dotted blue curve represents the brain postweaning. Each curve is surrounded by a shaded area indicating the 95% confidence interval. P values for group differences: TT: 6.7e-5, Lmean: 5.4e-5, Lmax: 2.4e-3, Lentr: 1.0e-4, DET: 7.9e-5, and RR: 1.5e-4. All P values are derived from the Wilcoxon test using frequencies from 0 to 70 Hz. This figure was created using the R statistical computing package.⁴⁹ Curves are averages across all patients in respective groups.

FIG. 3.

Percentage change in the area under the multiscale curves when a patient is weaned off of antiseizure medication. This figure was created using the R statistical computing package.⁴⁹

Effects of Antiseizure Medication on the Seizure Onset Zone

For each patient, we determined which electrodes were closest to the seizure onset zone (SOZ), the area of the brain where seizures start,³³ by locating the respective electrodes involved in seizure onset. For this subanalysis, we excluded patients with generalized seizures. This allowed for comparison of medication effects on the SOZ versus non-SOZ part of the brain (n = 65). Figure 4 shows heat maps on the standard 10 to 20 montage for each nonlinear measure from one patient. The relative size of the medication effect on RQA measures differs between the SOZ and non-SOZ regions of the brain. TT, Lmean, and Lentr show a larger effect on the non-SOZ region while Lmax, DET, and RR show a larger effect on the SOZ region. Regardless of the direction of change, the SOZ and non-SOZ portions of the brain respond differently to medication.

FIG. 4.

Heat maps of antiseizure medication effect size. Heat maps on the 10 to 20 system illustrating the relative change in each nonlinear dynamical value across all 67 patients, when given medication. Because seizure onset zone varies across each patient, the shown location is illustrative. We can observe that the size of the medication effect differs between the seizure onset zone and other regions of the brain.

Confounding Factors Related to Long-Term Monitoring Patients

LTM patients are presurgical patients with refractory epilepsy. Thus, this heterogeneous cohort includes patients with various seizure frequencies, treatment regimes, and responses to medication. To control for these factors, we stratify patients by (1) expected seizure frequency, (2) number of medications, and (3) a clinical chart review of “historical response to medication.”

(1) Expected Seizure Frequency

We used a measure of “expected seizures per day” computed from clinical record data as the ratio of seizure frequency (daily, weekly, monthly, and annually) to the number of seizures that occurred in the past 30 days. Stratifying patients with this information (n = 50) into 4 groups based on the in-cohort quartile of their expected seizures per day (n = 12, 8, 18, and 12, respectively), we observed that patients in the second and third quartile (i.e., with a moderate number of expected seizures per day) had the largest measured medication effect. Figure 5 shows the corresponding multifrequency curves (statistical values are presented in Table 2), and Figure 6 illustrates the change in the area under the curve for each quartile and each nonlinear measure. TT continued to show the largest medication effect, followed by Lmean, Lmax, Lentr, DET, and RR.

FIG. 5.

Multifrequency graphs for patients premedication and postmedication weaning, controlling for seizure burden. The multifrequency graphs are stratified by quartiles of expected seizures per day. The multifrequency curves represent trapping time (TT), Lyapunov max (Lmax), mean (Lmean), entropy (Lentr), determinism (DET), and recurrence rate (RR). The x-axis is the frequency, and the y-axis is the respective value. The red curves represent the electrodynamics of the brain for patients with the lowest expected number of seizures per day, followed by green, blue, and then purple. The purple curves represent patients with the highest number of expected seizures. The solid curves represent the EEG intervals where patients are on a high amount of medication, and the dotted curves represent the EEG intervals postweaning. Each curve represents the average for all patients in the group and is surrounded by a shaded area indicating the 95% confidence interval. Table 2 presents quantitative statistics to accompany these curves. This figure was created using the R statistical computing package.⁴⁹

TABLE 2.

P values are Related to the Curves in Figure 5

Nonlinear Measure	Drug High vs. Low
	1st Quartile	2nd Quartile	3rd Quartile	4th Quartile
TT	7.3e-6	1.1e-11	7.8e-3	5.8e-5
Lmean	9.4e-7	1.2e-11	1.1e-3	2.0e-4
Lmax	3.6e-3	9.9e-4	1.4e-3	1.7e-2
Lentr	7.3e-4	2.4e-9	9.9e-3	4.9e-4
Determinism	2.4e-5	6.2e-10	7.2e-3	4.6e-5
RR	3.4e-4	5.4e-10	3.2e-3	2.3e-3

FIG. 6.

Percentage change in the area under the multiscale curves when a patient is weaned off of antiseizure medication, stratified by a patient's expected seizures per day. This figure was created using the R statistical computing package.⁴⁹

(2) Number of Medications

Patients in this heterogeneous population are treated by a variety of medications including monotherapy and polytherapy regimens. Grouping patients who are treated with monotherapy (n = 5), patients on two to three medications (n = 52), and patients on four or more medications (n = 4), we created multifrequency curves for each measure (Fig. 7; statistical values for these curves are presented in Table 3). All groups saw a difference in nonlinear measures from preweaning to postweaning; however, patients on four or more medications showed a distinct opposite effect on the nonlinear measures than patients on 1 medication and patients on 2 to 3 medications. Patients on four or more medications show a distinct medication effect by a decrease across all nonlinear measures, while other patients showed an increase.

FIG. 7.

Multifrequency graphs for patients premedication (dashed) and postmedication (solid) weaning, controlling for the number of medications in the treatment regimen. The multifrequency graphs are stratified by the variety of different medications in their treatment regimen. The multifrequency curves represent trapping time (TT), Lyapunov max (Lmax), mean (Lmean), entropy (Lentr), determinism (DET), and recurrence rate (RR). The x-axis is the frequency, and the y-axis is the respective value. The red curves represent the electrodynamics of the brain for patients on monotherapy. The green and blue curves represent patients on polytherapy. The green curve represents patients on 2 to 3 medications, and the blue curve represents patients on 4 or more medications. The solid curves represent the EEG intervals where patients are on a high amount of medication, and the dotted curves represent the EEG intervals postweaning. Each curve represents the average for all patients in the group and is surrounded by a shaded area indicating the 95% confidence interval. Table 3 presents statistical information for the curves in this figure. This figure was created using the R statistical computing package.⁴⁹

TABLE 3.

P values Shown Here are Associated With the Curves in Figure 7 and Represent the Significance of Group Differences Between High and Low Medication Levels for Each of the Three Categories (Monotherapy, 2–3 Meds, 4+ Meds)

Nonlinear Measure	Drug High vs. Low
	Monotherapy	2–3 Meds	4+ Meds
TT	0.2	9.4e-4	0.014
Lmean	0.0034	0.0039	0.0028
Lmax	9.0e-5	0.054	0.024
Lentr	0.56	6.2e-4	0.056
Determinism	0.05	0.0030	0.0073
RR	0.74	0.0037	1.9e-4

(3) Response to Medication

Patients were stratified based on their clinically determined, historically reported short-term response to medication into groups of patients with “reduction in seizure frequency” (n = 31) versus “no reduction in seizure frequency” (n = 25). The first group is defined by patients that show at least some improvement in seizure burden, as evidenced by the reported reduction in seizure frequency after medication initiation, and the second group is defined by patients who do not show any clinically measurable change in seizure frequency after medication initiation. Figure 8 shows the corresponding multifrequency curves, with statistical values that accompany these curves in Table 4. For most of the nonlinear measures (Lentr, Lmax, Lmean, and TT), the corresponding values increase in both groups when the patients are weaned off of medication. However, patients with no reduction in seizure frequency show higher multiscale curves across all nonlinear measures.

FIG. 8.

Multifrequency graphs for patients premedication and postmedication weaning, controlling for the clinically determined response to medication. The multifrequency graphs are stratified by a reported “reduction in seizure frequency” and “reduction in seizure frequency” on medication. The multifrequency curves represent trapping time (TT), Lyapunov max (Lmax), mean (Lmean), entropy (Lentr), determinism (DET), and recurrence rate (RR). The x-axis is the frequency, and the y-axis is the respective value. The red curves represent the electrodynamics of the brain for patients who show improvement, and the blue curves represent patients do not have the reported reduction in seizure frequency. The solid curves represent the EEG intervals where patients are on a high amount of medication, and the dotted curves represent the EEG intervals postweaning. Each curve represents the average for all patients in the group and is surrounded by a shaded area indicating the 95% confidence interval. Table 4 presents statistical information for the curves in this figure. This figure was created using the R statistical computing package.⁴⁹

TABLE 4.

P values for Group Differences Between High and Low Drug Levels, Comparing Changes in the Sensors Closest With the Seizure Onset Zone to Remaining Sensors

Nonlinear Measure	Drug High vs. Low
	Seizure Onset Zone	Rest of the Brain
TT	0.011	0.0021
Lmean	0.0022	0.0090
Lmax	0.0077	0.11
Lentr	0.0041	0.0085
Determinism	0.0014	0.019
RR	1.4e-4	0.17

This table accompanies Figure 8.

DISCUSSION

Our findings are consistent with our hypothesis that decreased complexity in the brain is an indication of increased epileptogenicity because medications are weaned from the system. This study does not consider a single “number” to characterize medication effects but considers a more complicated quantification based on several nonlinear measures of complexity across a frequency spectrum. We included power with our nonlinear measures, and it did not change with the medication level. The greatest differences were in the gamma frequency range, above 30 Hz. While higher frequencies are generally thought to be poorly resolved in scalp EEG and potentially affected by artifact, recent research demonstrates that high-frequency oscillations can be detected in scalp EEG, although these features are low amplitude and easily obscured by noise or artifact.^34,35

Seizures are believed to arise in the setting of neuronal imbalance on cellular or anatomical levels. When the inhibitory systems fail, this imbalance permits hyperexcitability of neural circuits, entrainment, and recruitment of other neurons, resulting in synchronous activity.^36,37 This seizure-inducing excitability may be measurable and reduced by ASMs.³⁸ All nonlinear measures computed from RQA decreased when patients were on high amounts of medication, compared with when the medications were weaned (Fig. 2). This aligns with the physiological interpretation of the nonlinear measures to some extent, although the mathematical description of some of these measures cannot yet be interpreted physiologically. More complex systems, as reflected in RQA values, are less likely to synchronize, reducing seizure likelihood. For example, lower Lentr during ASM-low may signify neuronal synchrony and a higher likelihood for seizures to occur. Similarly, DET is a measure of predictability, and higher DET is associated with greater neural synchrony, as seen in ASM-low intervals. Overall, we interpret higher levels of chaos as “healthier,” as suggested by earlier studies of physiological dynamics.^39,40 Each of these nonlinear dynamic interpretations aligns with our results, suggesting that nonlinear dynamical values may reflect changes in epileptogenicity brought on through medication effects.

When differentiating between seizure-onset areas, compared with nonseizure onset areas, we observed that these regions reacted differently to medication (Fig. 4). Although all nonlinear measures showed a medication effect, and a difference in medication effect size between SOZ and non-SOZ of the brain, not all nonlinear measures showed the same directional change. TT, Lmean, and Lentr increased more in the non-SOZ region of the brain than the SOZ while Lmax, DET, and RR increased more in the SOZ region than in the non-SOZ. This is not unexpected because the different nonlinear measures capture different aspects of brain dynamics and medication effects on dynamics. A possible explanation for this is that (1) TT, Lmean, and Lentr capture an overall larger medication effect in the whole brain than Lmax, DET, and RR and/or (2) TT, Lmean, and Lentr are more sensitive to unrelated changes occurring in the regions of the brain that most of these patients have as non-SOZ. The first explanation holds and can be observed in Figure 3.

Our subanalysis findings determined that the size of the medication effect may correlate with a patient's seizure frequency, the number of medications, and clinically determined response to medication. Confounding factors include the following.

(1) Seizure Frequency

When stratifying patients by a computed expected seizures per day metric, patients with moderately frequent seizures relative to the rest of the cohort had larger medication effects. Patients with relatively few seizures or who had the highest seizure frequency had a relatively smaller measured medication effect than those with moderate seizure frequency. We do not have a compelling explanation for these results that connects the nonlinear measures to neurophysiology. It is important to note that reported seizures per day are based on patient or family reports and related chart review.

(2) Monotherapy Versus Polytherapy

Grouping patients by the number of medications in their treatment regimens illustrated that patients on four or more medications had higher values across all nonlinear measures (Fig. 7) as compared with other patients with fewer medications, possibly indicating higher medication effect and higher epileptogenicity. Patients on monotherapy had the lowest multiscale values at ASM-high. The higher epileptogenicity in patients on four or more ASMs may explain their need for a large number of medications. The ASM-high state showed higher epileptogenicity than when the patients were weaned off of medication. There could be many potential reasons for this opposite effect. For instance, some medications, such as certain psychostimulants and antibiotics, may decrease the seizure threshold and thus epileptogenicity,⁴¹ and some ASMs may precipitate seizures or alter their semiology.⁴² It is noteworthy that some antiseizure drugs may interact with others, possibly leading to potentiation or decrease of their cumulative effect.⁴³ Furthermore, medications may have persistent effects even after weaning, especially those with longer half-lives, and may vary by patient. These medication effects may extend into our ASM-low EEG segments. Future studies that focus on drug levels or half-lives could address this potentially confounding effect.

(3) Historical Response to Medication

We stratified patients by their clinically determined response-to-medication. For most nonlinear measures, patients in the “no reduction in seizure frequency” group had the highest nonlinear values postweaning, consistent with higher epileptogenicity, and the “reduction in seizure frequency” group preweaning had the lowest epileptogenicity parameters (Fig. 8). The patients who had been clinically determined to have no improvement on medication did show some medication effect, presenting as lower nonlinear measures preweaning compared with postweaning, indicating some increase in epileptogenicity when weaned off of medications. However, this measured electrodynamic effect in the brain was not enough to result in therapeutic changes.

Challenges

Findings from our cohort of presurgical patients that differ in their age, sex, diagnoses, comorbidities, MRI findings, and seizure burden may not be generalizable to a wider population of patients with seizures. Epilepsy severity may differ among our patients, which may interfere with the consistency of our results, because different patient characteristics may generate different epileptogenicity profiles and multiscale values. Although we examined several features that may relate to epilepsy severity (seizure frequency, number of medications, seizure onset zone, and response to medication), future studies to further account for indicators of epilepsy severity are warranted. We did not account for sleep–awake states, which may affect the EEG and RQA values. Our measure of seizure frequency was based on patient or family reports; this may be confounded by underreporting or overreporting and limitations of chart documentation related to seizure frequency. Medications may have persistent effects even after weaning, especially those with longer half-lives. These medication effects may extend into our ASM-low EEG segments.

Future studies that focus on drug levels or half-lives could address this potentially confounding effect.

In this study, we did not account for interictal findings or epileptogenic zone in the analyzed EEG excerpt. Similar nonlinear studies were able to identify differences in the EEGs of patients with epilepsy and the EEGs of controls with normal EEGs, with the absence of seizures, spikes, or other epileptiform activity.²³ This computational approach to measuring medication effects may therefore identify changes in epileptogenicity independently of seizures, seizure types, or EEG changes such as epileptiform discharges and high-frequency oscillations. Nevertheless, patients with presence or absence of interictal findings, such as high-frequency oscillations, spikes, or slowing, and amount of interictal epileptiform activity may show a different response to medication, and epileptogenic potential may be a confounder of related measures and is an important aspect to address in the future. Many patients were treated with benzodiazepines, which are known to affect EEG power in the beta frequency range.⁴⁴ We did not review these recordings for artifacts, and note that muscle artifacts have been shown to affect high-frequency EEG activity, but not MEG.^45,46 Higher frequency oscillations have been detected from surface EEG,⁴⁷ and this should be considered in future studies.

Although the results presented here are a promising first step toward a digital biomarker to monitor antiseizure medication effects, larger cohorts, with individual accuracies rather than just group comparisons, will be needed for analytical and clinical validation.

CONCLUSION

Multifrequency nonlinear values derived from EEGs seem to correlate with medication effects in the brain. Our results suggest that it may also be possible to measure the therapeutic efficacy of medication on patients with varying seizure frequencies, on patients with monotherapy or polytherapy treatment regimes, on specific targeted regions of the brain, and on patients who are generally considered to have not improved on medication. The multifaceted underpinnings of epilepsy have recently led to a change in perspective of the condition as a spectrum disorder.⁴⁸ With a broader definition, more medications will enter the space and better metrics to measure medication efficacy are crucial. This pilot study takes the first step toward guiding clinical decisions on medication choices through quantitative measures of medication efficacy.

Supplementary Material

jcnp-41-72-s001.docx ^{(18.2KB, docx)}