Epilepsy is Heterogeneous in Early-Life Tuberous Sclerosis Complex

Summary:

Objective:

Epilepsy in Tuberous Sclerosis Complex (TSC) typically presents with early onset, multiple seizure types, and intractability. However, variability is observed amongst individuals. Here, detailed individual data on seizure characteristics collected prospectively during early life were used to define epilepsy profiles in this population.

Methods:

Children ages 0–36 months were followed longitudinally. Caregivers kept daily seizure diaries, including onset and daily counts for each seizure type. Patients with >70% seizure diary completion and >365 diary days were included. Developmental outcomes at 36 months were compared between subgroups.

Results:

Epilepsy was seen in 124 of 156 (79%) participants. Seizure onset occurred from 0 to 29.5 months; 93% had onset prior to age 12 months. Focal seizures and epileptic spasms were most common. Number of seizures (for median 897 days) ranged from 1 to 9,128. Hierarchical clustering based on six metrics of seizure burden (age of onset, total seizures, ratio of seizure days to non-seizure days, seizures per seizure day, and worst 7- and 30-day stretches) revealed two distinct groups with broadly favorable and unfavorable epilepsy profiles. Subpopulations within each group showed clinically meaningful differences in seizure burden. Groups with higher seizure burden had worse developmental outcomes at 36 months.

Significance:

Although epilepsy is highly prevalent in TSC, not all young children with TSC have the same epilepsy profile. At least two phenotypic subpopulations are discernible based on seizure burden. Early and aggressive treatments for epilepsy in TSC may be best leveraged by targeting specific subgroups based on phenotype severity.

Introduction:

Tuberous Sclerosis Complex (TSC) is an autosomal dominant, systemic disease affecting ~1 in 6,000–10,000 live births ¹. In over two-thirds of cases, the disease is sporadic. There are two causative genes, TSC1 and TSC2 ^2–4. The core pathophysiology of TSC is dysregulated cellular growth, which leads to the proliferation of hamartomas in multiple body systems, including the brain ⁵.

Neurological manifestations of TSC account for much of the clinical morbidity and medical utilization associated with the disease. Autism features and intellectual disabilities are seen in >50% of patients ⁶. Epilepsy is even more problematic, affecting 80–90% of patients ⁷. Early epilepsy onset is common and is associated with high rates of medical intractability and long-term neurodevelopmental deficits ^8–10. Epileptic spasms, which affect ~30% of TSC patients, may also be an independent predictor of worse developmental outcomes ^{7, 11, 12}. In general, a TSC2 mutation confers a worse epilepsy phenotype than a TSC1 mutation ^13–15, even in mice ¹⁶. Nonetheless, variability in epilepsy severity has been noted within patients of the same genotype ¹⁷, and not all TSC patients with epileptic spasms have poor outcomes. At least one previous study has specifically reported a handful of TSC patients (4.4% of their cohort) who only experienced one seizure ¹⁸. Despite these observations, little work has been done to describe the individual heterogeneity of epilepsy in TSC with regard to day-to-day variability in seizure frequency and seizure type, particularly during the critical early years of brain development.

Understanding individual variability in epilepsy phenotype in TSC is more than just descriptive. Strategies are being developed for identifying TSC patients at greatest risk for developing epilepsy before they experience their first clinical seizure ^19–21, paving the way for clinical trials evaluating early, preventative treatment of epilepsy in TSC with vigabatrin ^{22, 23}. Early treatment with mTOR inhibitors is also being explored as a preventative anti-epilepsy therapy in this high-risk population ²⁴. Stratifying by epilepsy phenotype in TSC, especially in the first years of life, is crucial for optimal application of such precision therapies.

Here we provide detailed characterization of individual seizure onset and daily frequency patterns collected prospectively in a large multicenter longitudinal study of TSC patients starting as early as birth and continuing through 36 months of age. Machine learning techniques were applied to identify subpopulations within this age group based on individual patterns of seizure burden. Resulting subgroups were examined for differences in developmental outcomes.

Methods:

Subject recruitment

Infants for this analysis were drawn from two NIH clinical trials: Early Biomarkers of Autism in Infants with Tuberous Sclerosis Complex (TSC) (clinicaltrials.gov, NCT01780441) and Potential EEG Biomarkers and Antiepileptogenic Strategies for Epilepsy in TSC (clinicaltrials.gov, NCT01767779). Enrollment occurred from September 2012 to December 2018. Five U.S. sites, comprising the Tuberous Sclerosis Complex Autism Center of Excellence Research Network (TACERN) participated in these studies (Cincinnati Children’s Hospital Medical Center, Boston Children’s Hospital, University of Alabama at Birmingham Medical Center, University of California at Los Angeles and McGovern Medical School at the University of Texas Health Science Center at Houston). IRB approval was obtained at each site, and informed consent was obtained from each family before study participation.

Inclusion criteria for both studies required meeting clinical or genetic criteria for definitive diagnosis of TSC ²⁵. Patients also could not have significant prematurity or other perinatal complications that would independently confer additional risk of epilepsy. NCT01780441 participants were age 3–12 months at enrollment whereas NCT01767779 participants enrolled between 0–6 months. Seizure data, regardless of time of enrollment, was collected starting from birth, although daily seizure diary use did not start until actual enrollment. Participants in NCT01780441 were eligible even if epilepsy was diagnosed prior to enrollment. NCT0176779 did not allow enrollment for those with prior seizures or anti-seizure treatment. Patients with subependymal giant cell astrocytoma (SEGA) or previously treated with mTOR inhibitors were also excluded.

Study design

Each patient was evaluated longitudinally, starting at enrollment and continued with follow-up at chronological ages 3, 6, 9, 12, 18, 24 and 36 months. At each visit, clinical data that included basic demographics, medical and family history, developmental therapies (type and frequency), co-morbid conditions and medications were collected, and physical examinations were conducted. Developmental assessments were also conducted. Testing was performed by a licensed psychologist and/or speech therapist blinded to each participant’s clinical history. The full battery of tests performed has been previously described ⁸. The current developmental analysis was limited to each patient’s last result from the Mullen Scales of Early Learning (MSEL) ²⁶, which is a proxy for global development. For most patients (n=83/88), the last MSEL was administered at 36 months; the last MSEL was at 24 months for 4 patients and at 18 months for 1 patient. To avoid the confound of floor effects in this cognitively impaired population, we computed MSEL developmental quotients (DQs) rather than MSEL standardized scores for each participant ⁸. Specifically, the ratio of each participant’s raw age equivalent score to their chronological age was determined for each MSEL subdomain, which were then averaged (except gross motor, as it is not used to calculate the Early Learning Composite Score for the MSEL) to determine the overall DQ.

Patient diaries were used for seizure tracking, and caregivers were instructed to record every occurrence of individual seizure type(s) each day. To ensure the current analysis was representative of seizure patterns over time, we established a minimum threshold for seizure diary compliance. Each included participant had to have at least 70% of expected seizure diary days filled out, and the duration of seizure diary days recorded had to span a minimum of 365 days.

To optimize seizure recognition, caregivers were shown an educational DVD demonstrating multiple examples of seizure types frequently encountered during infancy in TSC, including epileptic spasms and focal seizures with or without impairment of consciousness ^{20, 27}. Caregivers identified seizure types by descriptive terms. Semiological classification (e.g., focal seizure without impairment of consciousness) was then assigned during study visits by the treating pediatric neurologist using ILAE criteria, except that epileptic spasms were not distinguished as focal or generalized. Discrete focal and generalized seizures were quantified individually for each occurrence by type. For epileptic spasms, there was variability between families as to how seizures were quantified. Some caregivers counted spasm clusters as a single event, whereas others counted each individual spasm. We therefore elected not to infer counts at either the individual spasm or cluster level in the analyses. Rather, we coded any day with any spasms as a “1” for each participant, and any day with no spasms as a “0”.

Statistical analysis

We identified the following eight measures as metrics of seizure burden (i.e., epilepsy severity): 1) total number of seizures reported (normalized by total number of included study days to account for variable lengths of enrollment; i.e., overall seizure frequency); 2) seizures per seizure day (i.e., number of seizures on a day that had any seizures); 3) proportion of total days that were seizure days (where a ‘seizure day’ is any day with at least one seizure); 4) peak seizure frequency by week (rolling 7 day period), in terms of total number of seizures; 5) peak seizure frequency by month (rolling 30 day period), in terms of total number of seizures; 6) peak seizure frequency by week, in terms of proportion of days that were seizure days; 7) peak seizure frequency by month, in terms of proportion of days that were seizure days; and 8) age of onset of epilepsy.

Hierarchical clustering (a commonly used unsupervised machine learning approach) was used to ascertain natural groupings of patients based on epilepsy profiles within the cohort. The eight metrics of seizure burden described above were extracted for each participant. Each raw metric was Z-score transformed to normalize skews in the distributions within each metric and to account for variable scales between metrics. Recognizing the necessary relatedness of some of these metrics but wishing to preserve as much unique information as possible, we formally assessed for collinearity between the eight variables and removed those with > 0.80 collinearity. The hierarchical clustering function in Matlab was applied on the remaining metrics using the Euclidean distance method and the ward linkage criterion ²⁸. We then examined the distributions of the metrics of seizure burden in the resulting patient sub-groups using standard descriptive statistics, as well as scatter plots to aid in the visualization of the distributions. We acknowledge the circularity in this approach (describing our groups based on the features that were used to define them). However, this kind of circularity is not problematic since the purpose of its application here is to describe the data rather than analyze it.

We additionally queried for between-group differences in external variables not used to identify the subgroups, i.e., variables about which the hierarchical clustering was agnostic. These variables included seizure types (prevalence of focal seizures, epileptic spasms, or both) and developmental outcome (DQ). Unpaired t-tests (with alpha set at 0.05) were used to assess for the statistical significance of group differences.

Results:

Patient and Seizure Diary Characteristics

One hundred fifty-six patients were enrolled (Figure 1). TSC2 mutations and variants (103/156; 66%) were more prevalent than TSC1 mutations and variants (17/156; 11%), consistent with prevalence rates reported for the TSC population generally ^{13, 14} (Table 1). Epilepsy occurred in 79% (n=124/156) overall, with rates higher in those with TSC2 mutations and variants (86%) than those with TSC1 mutations and variants (35%). Focal seizures (97/156 = 62%) and epileptic spasms (89/156 = 57%) were most common. Seizure onset ranged from 0 to 29.5 months, with most (93%) occurring before 12 months (median 4.7 months). Seizure diary recording began upon enrollment, and average diary compliance was high overall, at 85.5% (diary completion days/total diary days). Thirty-three participants (21%) were excluded from additional analysis due to diary compliance <70% (n=27) or recording duration less than 1 year (n=6) (Figure 1). Thirty-five of the remaining 123 patients (28%) did not report any seizures during the duration of the study, although 8 of these 35 reported a history of seizures prior to enrollment. Thus, 88 of the 123 potentially eligible patients (72%) reported at least one seizure during the study and were the focus population used for assessing individual epilepsy patterns of early TSC.

Figure 1:

CONSORT-like flow diagram depicting patient selection

Table 1:

Subject characteristics

	Sex	Genetic Status	Age at Enrollment (months) (± S.D.)	Age at Completion (months) (± S.D.)	Seizure onset before enrollment (%)
All participants (n=156)	78F 78M	17 TSC1 103 TSC2 14 NMI 22 Unknown	5.6 ± 3.2	35.3 ± 6.2	47%
Met inclusion criteria, reported no seizures during recording period (n=35)	16F 19M	11 TSC1 19 TSC2 3 NMI 2 Unknown	5.0 ± 3.0	35.7 ± 4.5	23%
Met inclusion criteria, reported at least one seizure during the recording period (n=88)	43F 45M	27 TSC1 66 TSC2 10 NMI 10 Unknown	5.8 ± 3.3	36.5 ± 2.5	53%

Overall, 78,951 patient diary days and 75,087 separate seizures were analyzed from these 88 patients (Supplemental Table 1). Median reporting duration was 897 days, or ~2.5 years (range 498–1261 days). Individually, seizure frequency was highly variable. Total seizures ranged from 1 to 9,218. Notably, 32/88 patients (36%) reported 10 or fewer seizures for the entire duration of enrollment. Per patient, days with at least one seizure ranged from 1–803, representing 0.1–90.0% of total diary days recorded. Other seizure burden metrics, including number of seizures per seizure day, peak seizure frequency by week or month, and age of epilepsy onset demonstrated similarly high variability in this cohort. Many patients (n = 15/88) had been diagnosed with epileptic spasms prior to study enrollment and had been successfully treated. As a result, diaries captured focal seizures in more patients (n = 72/88; 82%) than epileptic spasms (n = 48/88; 55%), although the overall prevalence of epileptic spasms in our study cohort was 72% (n = 63/88). Primarily generalized seizures other than spasms were relatively uncommon (15%).

Hierarchical clustering for detection of groups with similar seizure burden

Two metrics of seizure burden (number of seizures per month and number of seizure days per week) were removed for redundancy (collinearity >0.80) with the remaining metrics. The remaining six metrics (total number of seizures, proportion of days with at least one seizure, number of seizures per seizure day, peak number of seizures per week, peak proportion of seizure days per month, and age of seizure onset), calculated for each of 88 patients, yielded a hierarchical clustering solution with cophenetic R of 0.5782. Figure 2 demonstrates the individual variability of just one of the seizure burden metrics used (proportion of seizure days/total days) in each 7-day interval throughout the duration of the study, sorted according to this hierarchical clustering solution.

Figure 2: Heat map of individual seizure burden over time measured by the number of seizure days each week.

Patients are organized along the y-axis according to hierarchical cluster membership. Each square across the x-axis represents one week of study enrollment starting at birth through study completion (0–36 months). Each row represents one individual patient (n=88) who reported at least one seizure during the reporting period. Squares are color coded according to the number of seizure days (0–7) for each week that each patient was enrolled and data was complete. Blank squares indicate missing data or time outside of (before or after) the reporting period.

The dendrogram shown in Figure 3A graphically depicts the clustering solution. Two threshold choices for dissimilarity are indicated by dotted lines that cross the y-axis of the dendrogram. The cohort segregated into two large populations with generally favorable (n = 43) and unfavorable (n = 45) profiles with regard to seizure burden when a high dissimilarity distance threshold was applied. The median values for all seizure frequency metrics were lower for the ‘favorable’ group compared to the ‘unfavorable’ group, and the median participant age at seizure onset was older for the ‘favorable’ group compared to the ‘unfavorable group’. At a lower dissimilarity distance threshold, the ‘favorable’ and ‘unfavorable’ groups further separated into five sub-groups with additional distinct clinical profiles. Two of these sub-groups were nested within the ‘favorable’ group, and three were nested within the ‘unfavorable’ group; each comprised 9–30 participants. These seizure burden profiles defining these sub-groups are detailed in Figure 3B-D.

Figure 3: Characteristics of TSC subgroups according to seizure burden.

(A) Dendrogram depicting the hierarchical clustering solution for seizure burden in early TSC. Each of 88 participants is represented by a single tick on the x-axis. Dissimilarity distance is shown on the y-axis, with high and low thresholds applied to identify separable subpopulations; thresholds are depicted as dotted lines. (B) Median values for each seizure burden metric for each of the five sub-groups, by group. (C) Patterns for each seizure burden metric for each of the five sub-groups, by individual. Each is color-coded and organized according to cluster membership, starting with cluster A on the left and then showing clusters B, C, D and E. (D) Summary of seizure burden characteristics for each of the five sub-groups.

Both ‘favorable’ sub-groups (Clusters A and B) experienced relatively infrequent seizures. However, the groups differed in both age of seizure onset (median 5.2 vs. 12.5 months, respectively) and peak number of seizures days per week (5/7 vs 2/7 days, respectively). The three ‘unfavorable’ sub-groups (Clusters C, D, and E) tended to seize earlier and more often overall but differed from one another in a step-wise fashion with regard to average number of seizures per day (0.49 vs. 1.39 vs. 4.67, respectively). Furthermore, patients in Clusters D and E had similar and remarkably high proportions of days that were seizure days (56% and 62%, respectively), higher than what was seen in Cluster C (11%) and much higher than what was seen in the ‘favorable’ sub-groups (1% and > 0.0% for clusters A and B, respectively).

Cluster E was clearly the subgroup with the worst seizure frequency metrics in all categories and the youngest age of seizure onset. The profile of other metrics of seizure burden for the other two ‘unfavorable’ subgroups (Clusters C and D) were mixed, with one (Cluster C) characterized by fewer total seizures and greater seizure-free days but a younger age of onset and a greater propensity to have more days with multiple seizures in the same day.

Clinical correlation of seizure burden profiles with seizure types

Seizure types differed among sub-groups (Figure 4). In terms of how many participants in each cluster had ever been diagnosed with focal seizures, epileptic spasms, or both seizure types (concurrent or sequential; prior to enrollment or during the study period), 100% of the patients in Clusters D and E had focal seizures compared to only 77% in Clusters A-C. Cluster B patients were much less likely to have experienced epileptic spasms (31%) compared all other clusters (73–89%). The combination of focal seizures plus spasms was seen in the vast majority of patients in Clusters D and E (79% and 89%, respectively); was seen in about half the patients in Clusters A and C (53% and 55%, respectively), and was seen in only a small minority (15%) of patients in Cluster B.

Figure 4: Differences in seizure types among TSC subgroups separated by seizure burden.

Proportion of patients reporting focal seizures only, epileptic spasms only, or both focal seizures and epileptic spasms is shown for each of the five sub-groups, by group.

Clinical correlation of seizure burden profiles with developmental outcome

Seizure burden profiles in the ‘favorable’ group (Clusters A and B), were associated with higher DQs at study completion (median 77.3 and 83.1, respectively) (Figure 5). DQs were highest for Cluster B, the subgroup with latest seizure onset, lowest prevalence of epileptic spasms with or without focal seizures, and lowest seizure frequency overall. In contrast, DQs for all subgroups in the ‘unfavorable’ group (Clusters C, D, and E) were significantly lower (60.9, 45.7, and 56.4, respectively).

Figure 5: Differences in developmental outcome among TSC subgroups separated by seizure burden.

Box and whisker plots depicting the median developmental quotient (DQ), by group, at 36 months of age. Solid brackets indicate a significant pairwise group difference at p<0.05; dotted brackets indicate a significant pairwise group difference at p<0.005.

Discussion:

Epilepsy is very common in TSC, with the vast majority of affected individuals having seizures with onset in the first year of life. While phenotypic variability has long been appreciated in TSC generally, our study extends this understanding to identifiable patterns of seizure burden specifically. Based on daily seizure diaries collected in TSC infants 0–36 months of age followed longitudinally, we were able to show the extent of seizure burden heterogeneity in early-life TSC. A history of any epilepsy is a well-established risk factor for intellectual disability and neurodevelopmental disorders in TSC ^{7, 29, 30}. More recently, we and others have shown age at the time of seizure onset is also a key predictor of these risks ^{8, 9, 18}. The current study shows that various approaches to quantification of seizure frequency and seizure days, as metrics of seizure burden, can also be used to further refine these risk prediction models.

We applied machine learning to analyze >75,000 individual seizures occurring in nearly 90 infants during the 3-year monitoring period to identify within the overall group distinct sub-groups with clinical implications. We concurrently examined several complementary but non-redundant metrics of seizure burden in order to capture as much of the clinically meaningful epilepsy phenotype as possible. Hierarchical clustering revealed the naturally occurring patterns of similarity and dissimilarity in our large, multi-center cohort. Similar approaches have been adopted to identify meaningful patterns in complex data sets in epilepsy ³¹ and other areas of neurology ³². In TSC, machine learning has also been used to guide expectant management and timing of future evaluations to surveil for emergence and/or progression of overall disease based on the extent of multi-organ involvement ³³. As another example, convolutional neural networks have been utilized to identify and segment tubers on structural MRI^{34, 35}. In the current study we identified five sub-groups which differed on two features about which the clustering was agnostic. Qualitatively, they differed with regard to the prevalence of focal seizures, epileptic spasms, or the combination of those seizure types.

Quantitatively, they differed with regard to the 36-month DQs in the expected direction, with an inverse relationship between DQ and seizure burden. These observations are critical because they bolster the validity of the chosen hierarchical clustering approach. They also suggest the value of future extrapolations of this methodology. Specifically, it may eventually be possible to leverage data such as these to build prospective models predicting phenotypic sub-groups within early-life TSC. These predictions could then be used for guiding therapeutic decision-making, such as the selection of patients for specifically timed disease-modifying therapies, or the identification of patients requiring increased surveillance andintervention for cognitive or neurodevelopmental disorders ³⁶. For example, current epilepsy prevention strategies focus on early detection of EEG abnormalities and pre-symptomatic initiation of treatment with vigabatrin for all infants with TSC ^{19, 23}. If epilepsy severity were able to be stratified prospectively, some sub-groups of TSC patients predicted to have infrequent seizures may be able to avoid prophylactic treatment and the associated potential side effects. Other subgroups might be identified as likely to show particular benefit from early treatment with mTOR inhibitors, another increasingly common early treatment strategy³⁷ with potential side effects ³⁸, or could be targeted for earlier utilization of surgical epilepsy treatment ^39–41. In sum, the current data support the general concept that cohort stratification using machine learning techniques has the potential to evolve into an extremely valuable clinical tool, although further work is needed.

In an earlier analysis of this cohort using repeated measures mixed logistic modeling, we found that a younger age at time of seizure onset was most predictive of later developmental outcome ⁸. The present study’s hierarchical clustering solution again identified that the most favorable outcomes with regard to long-term seizure control and global developmental outcome were seen in the sub-group with the oldest median age at seizure onset (Cluster B, 12.5 months), whereas the least favorable outcomes were seen in the sub-group with the youngest age at seizure onset (Cluster E, 2.7 months). The remaining subgroups were not readily distinguishable, however, by median age at seizure onset that was typical for TSC ³³ (Clusters A, C and D, with ages at seizure onset of 5.2, 4.2 and 5.5 months, respectively), but differed significantly with regard to developmental outcome. The hierarchical clustering solution, using the additional metrics of seizure burden, provided an increased ability to differentiate populations of TSC infants with epilepsy when age of seizure onset alone is insufficient for risk prediction. On a practical level, these findings also yield the very hopeful insight that initial seizures occurring early in life need not seal one’s fate in TSC and that favorable outcomes can be obtained in infants if overall seizure frequency remains low or readily responds to treatment without subsequent recurrence. This seems to be true whether the seizures were focal-onset or epileptic spasms. Furthermore, those who were able to accumulate more seizure-free days over time appeared to do better, even if seizure frequency was worse at times compared to their peers. This emphasizes the continued need for adjusting anti-seizure treatment regimens and exploring new treatment options when current seizure control is unsatisfactory at these young ages, as any improvement may have incremental impact on long-term developmental outcome.

Our study does have limitations. Although we obtained excellent overall compliance with the daily diaries, we cannot know the extent to which caregivers may have been inaccurate in their seizure reporting (for example, see Elger and Hoppe⁴²). We also acknowledge that we had to make choices about which specific metrics of seizure burden to employ, and some authors may have made other reasonable choices ⁴³. Additionally, the present analysis did not address all potential covariates and confounds, including specific anti-seizure medications used, dosing, and duration of treatment. For example, we have shown that infants within this cohort with epileptic spasms who are treated with low dose vigabatrin are at increased risk for spasm recurrence ⁴⁴, demonstrating specific treatment decisions likely impact seizure burden. Correlation with structural disease burden and functional connectivity in the brain are also important and key for further understanding the biological mechanisms accounting for the phenotypic clustering that differentiates sub-groups within this population. To this end, brain MRI and EEG were collected prospectively in this cohort and analyses are ongoing at this time. A final caveat pertains to the specificity of the subgroups we identified. The five hierarchical clusters described herein need not represent immutable phenotypic categories within early-life TSC, although we did utilize a fairly large sample to obtain these results, and the groupings do appear to be valid. Replication of our specific findings in a distinct cohort of similarly acquired patients would be ideal, but for a rare disease such as TSC, this may be difficult to realize.

In conclusion, demonstration of high individual variability in epilepsy severity and burden in TSC is not new ^{7, 18}, but until now no large, prospective studies have provided the detail necessary to investigate patterns or identified clinically significant sub-groups within the overall population. With application of machine learning, we were able to identify distinct clinically relevant seizure burden profiles that are predictive of epilepsy and global developmental outcome.

Key bullet points:

• >78,000 diary days and >75,000 distinct seizures were prospectively recorded in infants (0–36 months) with Tuberous Sclerosis Complex.

• Seizure burden varied widely amongst patients.

• Hierarchical clustering, based on metrics of seizure burden, revealed sub-groups of patients with favorable and unfavorable phenotypes.

• Groups with more severe epilepsy phenotypes had worse developmental outcomes.