Accumulated seizure burden predicts neurodevelopmental outcome at 36 months of age in patients with tuberous sclerosis complex

Abstract

Objective

Epilepsy and intellectual disability are common in tuberous sclerosis complex (TSC). Although early life seizures and intellectual disability are known to be correlated in TSC, the differential effects of age at seizure onset and accumulated seizure burden on development remain unclear.

Methods

Daily seizure diaries, serial neurodevelopmental testing, and brain magnetic resonance imaging were analyzed for 129 TSC patients followed from 0 to 36 months. We used machine learning to identify subgroups of patients based on neurodevelopmental test scores at 36 months of age and assessed the stability of those subgroups at 12 months. We tested the ability of candidate biomarkers to predict 36‐month neurodevelopmental subgroup using univariable and multivariable logistic regression. Candidate biomarkers included age at seizure onset, accumulated seizure burden, tuber volume, sex, and earlier neurodevelopmental test scores.

Results

Patients clustered into two neurodevelopmental subgroups at 36 months of age, higher and lower scoring. Subgroup was mostly (75%) the same at 12 months. Significant univariable effects on subgroup were seen only for accumulated seizure burden (largest effect), earlier test scores, and tuber volume. Neither age at seizure onset nor sex significantly distinguished 36‐month subgroups, although for girls but not boys there was a significant effect of age at seizure onset. In the multivariable model, accumulated seizure burden and earlier test scores together predicted 36‐month neurodevelopmental group with 82% accuracy and an area under the curve of .86.

Significance

These results untangle the contributions of age at seizure onset and accumulated seizure burden to neurodevelopmental outcomes in young children with TSC. Accumulated seizure burden, rather than the age at seizure onset, most accurately predicts neurodevelopmental outcome at 36 months of age. These results emphasize the need to manage seizures aggressively during the first 3 years of life for patients with TSC, not only to promote seizure control but to optimize cognitive function.

Key points.

• At 36 months of age, children with TSC fall into two large neurodevelopmental subgroups, higher scoring and lower scoring.

• Subgroup at 36 months is 75% stable (compared to subgroup at 12 months).

• Three variables distinguish the two 36‐month subgroups: accumulated seizure burden (largest effect), 12‐month test scores, and tuber volume.

• Neither age at seizure onset nor sex distinguished 36‐month subgroups, although for girls there was a significant effect of age at seizure onset.

• Accumulated seizure burden and 12‐month test scores together predict 36‐month neurodevelopmental group with 82% accuracy and an AUC of .86.

1. INTRODUCTION

Tuberous sclerosis complex (TSC) is a genetic disorder affecting ~1 in 6000–10 000. ¹ A pathogenic variant in TSC1 or TSC2 ² , ³ , ⁴ disrupts mammalian target of rapamycin (mTOR) signaling, resulting in dysregulated cell growth and the development of hamartomas in multiple body systems. In the brain, these hamartomas are called tubers and are responsible for the neurologic manifestations of the disease. ⁵

Neurologic manifestations account for a significant portion of the morbidity in TSC. Most (80%–90%) patients develop epilepsy, often in the first year of life, ⁶ , ⁷ , ⁸ with a high rate of drug refractoriness. ⁹ , ¹⁰ Neurodevelopmental disorders including autism spectrum disorder and developmental delay are also common in TSC, affecting 30%–50% and 50%–60% of individuals, respectively. ¹¹ , ¹² , ¹³ , ¹⁴ , ¹⁵ Moreover, there is a well‐established association between seizures and development in early life. ¹⁶ , ¹⁷ , ¹⁸ , ¹⁹ In an interim analysis of the present data, we showed that the age at seizure onset was the most important factor predicting developmental delay at 24 months as measured by the Mullens Scales of Early Learning (MSEL), more so than the presence of infantile spasms or binned seizure frequency in the preceding 6 months. ⁷ Accumulated seizure burden was not evaluated in that study, and 36‐month developmental outcomes were not available for analysis.

This study extends the investigation of Capal et al. ⁷ by examining a later time point (36 months) and adding additional metrics, including a continuous variable capturing the accumulated number of seizure days. We supplement global developmental assessment by MSEL with measures of adaptive behavior as captured by the Vineland Adaptive Behavior Scales II (VABS‐II). The overarching objective was to clarify the relationship between epilepsy and development in early life TSC to provide actionable clinical guidance. We employed a three‐step approach: we characterized the neurodevelopmental profiles of children with TSC at age 36 months; we evaluated the stability of these profiles longitudinally; and we identified univariable and multivariable relationships between development at 36 months and other clinical factors.

2. MATERIALS AND METHODS

2.1. Subject recruitment

Participants derived from two National Institutes of Health clinical trials: (1) Early Biomarkers of Autism in Infants With Tuberous Sclerosis Complex (clinicaltrials.gov, NCT01780441) and (2) Potential EEG Biomarkers and Antiepileptogenic Strategies for Epilepsy in TSC (clinicaltrials.gov, NCT01767779). Enrollment began in September 2012 at five hospitals (Cincinnati Children's Hospital Medical Center; Boston Children's Hospital; University of Alabama of Birmingham Medical Center; University of California, Los Angeles; and McGovern Medical School at the University of Texas Health Science Center at Houston). Institutional review board approval was obtained at each site, and informed consent was acquired from all families. This research was conducted in accordance with the Code of Ethics of the World Medical Association (Declaration of Helsinki).

The primary inclusion criterion for both trials was a clinical or genetic diagnosis of TSC. ²⁰ Exclusion criteria included prematurity (<36 weeks estimated gestational age), perinatal complications, prior mTOR inhibitor use, subependymal giant cell astrocytoma (SEGA) requiring treatment, and prior epilepsy surgery. Participants were 0–12 months of age at enrollment.

2.2. Structure of prospective data collection

Patients were followed longitudinally with visits at baseline and ages 3, 6, 9, 12, 18, 24, and 36 months. Clinical information and physical examinations were obtained at each visit. This included names and dosages of all medications administered since the prior visit. Developmental assessments were administered as detailed previously. ⁷ Brain magnetic resonance imaging (MRI) scans were obtained as per guidelines for SEGA surveillance. ²¹ , ²² Electroencephalograms were performed regularly but not included here. Daily seizure diaries were recorded.

2.3. Neurodevelopmental assessments

We focused on the MSEL as a clinician‐administered assessment of global developmental functioning ²³ and the VABS‐II as a caregiver‐reported measure of behavior and adaptive functioning. ²⁴ The MSEL has five subscales (gross motor, fine motor, expressive language, receptive language, and visual reception) and the VABS‐II has 11 (play and leisure, interpersonal relationships, coping skills, receptive language, expressive language, written language, fine motor, gross motor, personal skills, domestic skills, community skills). For each participant for each subscale, we computed a developmental quotient (DQ) by dividing the age‐equivalent subscale score by chronological age and multiplying the result by 100. DQs reflect age‐corrected performance while avoiding floor effects. ⁶ , ⁷

2.4. Tuber volume calculation

Tuber segmentation was performed using the highest contrast T1‐weighted structural MRI scan at age >2 years. Images were automatically segmented ²⁵ then manually adjusted by a TSC imaging expert (J.M.P.), as described previously. ²⁶ Tuber distributions were registered individually to a common space using the Montreal Neurological Institute 1522009c nonlinear asymmetric template. Tuber volume was computed as percentage of cortical gray matter volume, akin to tuber brain proportion. ¹⁶

2.5. Accumulated seizure burden

Seizures were tracked using daily patient diaries. Prospectively, caregivers were shown an educational DVD demonstrating examples of seizure types frequently seen in infants with TSC, including infantile spasms and focal seizures. ²⁷ Caregivers were asked to record each occurrence of each seizure type(s) each day, with “types” identified in descriptive terms (e.g., “Type A = eyes deviate to left, stares off”). Semiological classification of “types” was then assigned during study visits by the treating clinician as per International League Against Epilepsy criteria (e.g., focal seizure without impairment of consciousness), except that epileptic spasms were not distinguished as focal or generalized. For spasms, some families counted spasm clusters as a single event, whereas others counted each individual spasm. Rather than infer counts, we coded any day with any spasms as a “1” for each participant and any day with no spasms as a “0.” To ensure that data were representative of individuals' seizure patterns over time, we established a minimum threshold for seizure diary compliance. Included participants were required to have at least 70% of expected seizure diary days filled out, and the duration of seizure diary days had to span at least 365 days. We have previously published a report characterizing seizure profiles in this population using this seizure diary methodology with these quality control criteria. ⁶

The objective of the present analyses was to capture a robust measure of each patient's longitudinal seizure burden. We choose a “seizure day” as a meaningful unit of seizure burden, where each day with at least one seizure of any type was counted as a “seizure day.” The number of seizure days accumulated over the entire study period was calculated for each patient for whom this was reported, including patients reporting zero seizures. We recomputed this sum considering separately the time from 0 to 12 months and from 12 to 36 months.

2.6. Antiseizure medication exposure

We calculated for each patient an antiseizure medication (ASM) count, defined as the sum of the number of ASMs reported at each study visit. Note that this is a proxy for cumulative ASM exposure and does not represent the number of distinct ASMs prescribed, because medications could be counted more than once. We also specifically noted which patients had ever been exposed to vigabatrin, hormonal therapy (prednisone or adrenocorticotropic hormone) and mTOR inhibitors (sirolimus or everolimus).

2.7. Statistical analysis

Descriptive and statistical analyses were performed using MATLAB, ²⁸ R, ²⁹ and Python. ³⁰

2.7.1. Hierarchical clustering at 36 months

Our first objective was to identify subgroups based on neurodevelopmental assessments at 36 months, reasoning that the last test scores would be the most predictive of future neurodevelopment. We employed hierarchical clustering to illuminate the underlying structure in the data. Input features were the neurodevelopmental subscale DQs from the MSEL and VABS‐II, Z‐score transformed. Not all patients were scored on all subscales at all time points; thus, we submitted to analysis only those subscales for which data were available for >90% of the patients. The gross motor subscale of the MSEL was omitted, as it is typically excluded from the MSEL composite. ²³ A cosine distance method and average linkage criterion were selected for clustering. ²⁸ We have used a similar approach to identify subgroups of patients based on seizure burden. ⁶

The resulting clustergram was visually inspected for subgroups. The distributions of the subscale scores within and between the resulting subgroups were queried using standard descriptive statistics. Unpaired t‐tests, corrected for multiple comparisons, were used to test for statistical differences between groups for each subscale score. This circular approach (assessing for significant differences in metrics between subgroups that were defined based on those metrics) was intentional, allowing us to assess whether each metric was different between subgroups or whether particular feature(s) drove the result.

2.7.2. Exploration of biomarkers associated with 36‐month subgroup

We next interrogated for between‐subgroup differences in variables not used in the clustering but that could instead serve as biomarkers of neurodevelopmental phenotype. These included prevalence of epilepsy, infantile spasms, and focal seizures; epilepsy surgery; sex; TSC pathogenic variant; tuber volume; age at seizure onset, accumulated seizure burden; and 12‐month DQs. Statistical tests included unpaired t‐tests (for continuous variables) and Fisher exact test and logistic regression (for categorical variables). Not all patients had all data points; each n is reported. Two‐sided p‐values are reported.

2.7.3. Hierarchical clustering at 12 months

We next asked whether neurodevelopmental subgroups were stable over time, specifically whether subgroups would shift if recomputed using data obtained at 12 months of age. Twelve months was chosen because testing is more reliable at that age compared to earlier ages and there were few missing data points at 12 months. Hierarchical clustering was thus repeated using the 12‐month Z‐score transformed subscale DQs. We identified which patients switched subgroups between time points. We additionally assessed the relationships between stability of subgroup and accumulated seizure burden, age at seizure onset, tuber volume, and 12‐month DQ.

2.7.4. Logistic regressions

Lastly, we tested candidate biomarkers for predicting 36‐month neurodevelopmental subgroup. Biomarkers, selected based on the results of the univariable analyses, included accumulated seizure burden, age at seizure onset, tuber volume, 12‐month DQ (combined across MSEL and VABS‐II), and ASM exposure. The distributions of each variable were evaluated and log‐transformed if skewed. A standard scaler was then applied so that between‐variable differences in scale would not influence the multivariable model. Univariable logistic regression analyses were computed for each variable individually. Lastly, a multivariable logistic regression model was computed starting with all five variables, with regularization applied (optimal parameter identified from a search space). Backward elimination based on the absolute values of the regression coefficients was used to identify the best multivariable model. Cross‐validation was achieved with 10 stratified k‐folds. Performance characteristics, including area under the curve, precision, recall, F1‐score, and overall accuracy, were evaluated for all models. Receiver operating characteristic curves were plotted.

3. RESULTS

3.1. Patient characteristics

Across the two clinical trials, n = 169 patients were enrolled. We included all patients who had MSEL and VABS‐II scores at the 12‐month and the final study visit (mean = 37 months of age, range = 34–44), yielding n = 129 patients. Patient characteristics are detailed in Table 1.

TABLE 1.

Patient characteristics and statistical test results.

	Full cohort	Higher scoring group	Lower scoring group	p	OR or t‐statistic
n	129	74	55
36‐month DQ, MSEL+VABS‐II overall, mean (SD)	73.1 (27.5)	92.9 (14.8)	46.5 (15.7)	<.0001	t = 17.11
36‐month DQ, VABS‐II, overall, mean (SD)	74.0 (30.5)	95.3 (19.3)	45.3 (15.7)	<.0001	t = 15.73
36‐month DQ, MSEL, overall, mean (SD)	72.4 (27.3)	90.6 (16.8)	47.9 (18.0)	<.0001	t = 13.83
36‐month DQ, MSEL verbal, mean (SD)	71.1 (28.9)	90.8 (17.9)	44.6 (17.3)	<.0001	t = 14.71
36‐month DQ, MSEL nonverbal, mean (SD)	73.9 (27.1)	90.6 (18.2)	51.5 (20.1)	<.0001	t = 11.54
n	129	74	55
12‐month DQ, MSEL+VABS‐II overall, mean (SD)	82.6 (23.9)	93.7 (18.7)	67.6 (22.0)	<.0001	t = 7.25
12‐month DQ, VABS‐II, overall, mean (SD)	82.7 (25.4)	93.5 (22.3)	68.2 (21.8)	<.0001	t = 6.42
12‐month DQ, MSEL, overall, mean (SD)	82.5 (25.1)	93.9 (18.3)	67.1 (24.9)	<.0001	t = 7.05
12‐month DQ, MSEL verbal, mean (SD)	73.7 (22.9)	83.7 (18.4)	60.2 (21.4)	<.0001	t = 6.69
12‐month DQ, MSEL nonverbal, mean (SD)	91.4 (29.8)	104.3 (21.5)	74.1 (31.0)	<.0001	t = 6.53
n	129	74	55
Epilepsy prevalence	78% (101/129)	64% (47/74)	98% (54/55)	<.0001	OR = .03
Infantile spasms prevalence	58% (75/129)	45% (33/74)	76% (42/55)	.0003	OR = .25
Focal seizures prevalence	65% (84/129)	45% (33/74)	93% (51/55)	<.0001	OR = .06
Epilepsy surgery during study period	20% (26/129)	13% (10/74)	29% (16/55)	.044	OR = .38
Sex, female	49.6% (64/129)	48.6% (36/74)	50.9% (28/55)	.8595	OR = .91
Exposure to vigabatrin, prevalence	71% (92/129)	54% (40/74)	95% (52/55)	<.0001	OR = .07
Exposure to mTOR inhibitor, prevalence	12% (15/129)	7% (5/74)	22% (10/45)	.0551	OR = .33
Exposure to hormonal therapy, prevalence	13% (17/129)	8% (6/74)	20% (11/55)	.0653	OR = .35
Number of ASMs prior to age 12 months, mean (SD)	2 (3)	2 (2)	3 (3)	.0027	t = 3.06
Number of ASMs between 12 and 36 months, mean (SD)	7 (6)	5 (4)	10 (5)	<.0001	t = 6.75
Number of ASMs over entire study period, mean (SD)	9 (7)	6 (6)	13 (7)	<.0001	t = 6.37
n	101	57	44
Proportion of TSC1 versus TSC2 pathogenic variants	13.9% (14/101)	19.3% (11/57)	6.8% (3/44)	.0871	OR = 3.27
n	107	60	47
Gray matter tuber volume, mean % (SD)	.0247 (.0216)	.0179 (.0171)	.0333 (.0236)	.0002	t = 3.89
n	101	47	54
Age at seizure onset, days, mean (SD)	172 (138)	199 (170)	149 (95)	.0656	t = 1.86
n	109	65	44
Accumulated number of seizure days from 12 to 36 months, mean (SD)	106 (180)	42 (124)	200 (206)	<.0001	t = 4.97

Note: Probability values shown are uncorrected.

Abbreviations: ASM, antiseizure medication; DQ, developmental quotient; MSEL, Mullens Scales of Early Learning; MSEL nonverbal, mean of MSEL visual reception and fine motor; MSEL verbal, mean of MSEL expressive language and receptive language; mTOR, mammalian target of rapamycin; OR, odds ratio; VABS‐II, Vineland Adaptive Behavior Scales II.

3.2. Clustering results

At 36 months, all subscales of the MSEL and VABS‐II had adequate data for clustering (>90% of patients) except for the VABS‐II writing subscale, which was excluded. The clustering solution at 36 months is depicted in Figure 1A as a clustergram, which includes a dendrogram and heatmap. One primary observation is the natural separation of the patients into two large subgroups, one with generally lower scores (n = 55) and one with generally higher scores (n = 74). Another observation is that there is some clustering of similar subscales along the x‐axis, for example, the expressive language subscores for the MSEL and VABS‐II. Unpaired t‐tests between subgroups for each subscale were all significant (p < .0001, Bonferroni‐corrected for multiple comparisons). Table 1 shows the group means and statistical results for the MSEL and VABS‐II composite scores as well as MSEL verbal (receptive language and expressive language) and nonverbal (fine motor and visual reception) domains.

FIGURE 1.

Hierarchical cluster result at (A) 36 months of age and (B) 12 months of age. Shown in panel A is the clustergram derived from hierarchical clustering of n = 129 patients using n = 14 developmental quotients (DQs) obtained at 36 months of age, n = 10 from the Vineland Adaptive Behavior Scales II (VABS) and n = 4 from the Mullen Scales of Early Learning (MSEL). Shown in panel B is the corresponding clustergram obtained using the n = 11 DQ subscales available at 12 months of age. In both cases, the clusters split nicely into two main subgroups. In each panel, each patient is represented as one tick on the colored dendrogram on the left, which splits into two large subgroups. Each individual patient is furthermore assigned one row across the page, with that patient's score for each subscale depicted as a shaded cell in one of the columns, labeled across the x‐axis (see abbreviation key below). Scores are Z‐score transformed DQs, as shown in the color scale, with warm colors positive and cool colors negative. Thus, the blue groups had generally lower scores, whereas the red groups had generally higher scores. Patients who clustered together on the dendrogram had similar developmental patterns, with the length of the line connecting two ticks proportional to the overall dissimilarity between two individuals. Similarly, the relative similarity or dissimilarity between the subscales can be inferred by examining the dendrogram at the top of the figure and noting the length of the line connecting any pair of columns. For example, in panel A, the shortest distance connects the MSEL visual reception and MSEL receptive language subscales, suggesting that among the 14 subscales, scores on the MSEL visual reception and receptive language subscales are the most congruent within individuals in this sample at 36 months of age. Comm, community; Cope, coping skills; Domes, domestic; ExpLan, expressive language; FinMo, fine motor; GroMo, gross motor; Pers, personal; Play, play and leisure time; RecLan, receptive language; Relat, interpersonal relationships; VisRec, visual receptive.

At 12 months, 11 subscales had adequate data for clustering; VABS‐II coping skills, domestic skills, and community skills subscales were omitted. Two large subgroups again resulted (Figure 1B). Most patients (n = 96/129, or 74%) fell into the same subgroup at 12 months as at 36 months, that is, were consistently higher or lower scoring.

Compared to those persistently higher scoring (n = 54), persistently lower scoring patients (n = 42) had larger tuber volumes (mean = .034% vs. .017%, t‐statistic = 4.28, p < .0001), earlier age at seizure onset (134 vs. 211 days, t‐test statistic = 2.7, p < .01), higher accumulated seizure burden (231 vs. 16 days, t‐statistic = 6.32, p < .0001), and lower 12‐month DQ (60 vs. 102, t‐statistic = 13.23, p < .0001). Those who changed subgroup in either direction tended to have tuber volumes, ages at seizure onset, accumulated seizure burdens, and 12‐month DQs toward the middle of the respective variable distributions (Figure 2).

FIGURE 2.

Accumulated seizure burden, age at seizure onset, tuber volume, and developmental quotient (DQ) at 12 months all varied as a function of neurodevelopmental subgroup stability. Patients either were persistently in the higher scoring subgroup (Hi/Hi; n = 54), were persistently in the lower scoring group (Lo/Lo; n = 42), or flipped between subgroups from 12 to 36 months (Flip; n = 33, including n = 13 who went from higher scoring to lower scoring and n = 20 who went from lower scoring to higher scoring). The boxplots show the distributions of accumulated seizure burden (in number of seizure days), age of seizure onset (in days), tuber volume (as percent of gray matter occupied by tuber), and DQ at 12 months (expressed as an average of the Vineland Adaptive Behavior Scales II [VABS] and Mullens Scales of Early Learning [MSEL] subtest scores at 12 months), each as a function of subgroup stability.

3.3. Relationship between subgroups and clustering‐naïve variables

The lower scoring subgroup had higher prevalence of epilepsy, infantile spasms, and focal seizures (Table 1). Only one child in the lower scoring subgroup did not have epilepsy, whereas more than one third of children in the higher scoring subgroup never seized. Focal seizures were also more prevalent in the lower scoring compared to the higher scoring subgroup. Consistent with these significant differences in seizure prevalence, patients in the lower scoring subgroup reported a significantly larger number of ASMs, whether counted prior to age 12 months, between 12 and 36 months, or over the duration of the entire study. Those in the lower scoring subgroup were significantly more likely to have been exposed to vigabatrin, with similar although nonsignificant trends for likelihood of having been exposed to hormonal therapy and mTOR inhibitors (Table 1).

Across the cohort, one fifth of patients underwent epilepsy surgery during enrollment. Patients in the lower scoring subgroup were more than twice as likely to have undergone surgery as those in the higher scoring subgroup (Table 1). Ninety percent (9/10) of those in the higher scoring subgroup experienced seizure reduction following surgery, compared to 69% (11/16) in the lower scoring subgroup (odds ratio = 4.09, p = .35).

The subgroups included roughly equal numbers of boys and girls, and there was no sex difference in the age at epilepsy onset (mean = 176 days for girls, 168 for boys, t‐statistic = .2993, p = .7654) or accumulated seizure burden from 12 to 36 months (mean = 116 seizure days for girls, 97 for boys, t‐statistic = −.5631, p = .5746).

Positive genetic test results (pathogenic variant or variant of unknown significance in TSC1 or TSC2) were available for n = 101 patients. The higher scoring subgroup had a higher proportion of TSC1 as compared to TSC2 pathogenic variants, although this trend did not reach significance (Table 1).

Tuber volumes were analyzed in n = 107 patients. Those in the lower scoring subgroup had a higher tuber volume by nearly twofold as compared to those in the higher scoring group (Table 1). Significant effects in the same direction were found when each lobe (combined across hemispheres) was considered separately (data not shown).

Seizures occurred either before or during the study in n = 101 patients. The lower scoring subgroup had a mean age at seizure onset nearly 2 months earlier than the higher scoring subgroup. This relationship trended toward significance (Table 1, Figure 3A). Post hoc analysis (Figure 3B) showed that the trend appeared to be driven by a differential effect in boys and girls. For girls (n = 53), later age at seizure onset was associated with being in the higher scoring 36‐month subgroup (t‐statistic = 2.06, p = .04), whereas for boys (n = 48) this effect was not present (t‐statistic = .34, p = .73). The interaction (sex × age at seizure onset) did not, however, reach significance by logistic regression (z = 1.075, p = .28).

FIGURE 3.

Age at seizure onset did not differ significantly between neurodevelopmental subgroups, although the effect was significant separately for girls but not boys. (A) Age of seizure onset for all patients with epilepsy (n = 101), stratified by neurodevelopmental subgroup at 36 months. There was a nonsignificant trend (p = .0656) for those in the higher scoring subgroup to have a later age at seizure onset. (B) Girls and boys separately. Asterisk indicates that for girls (n = 53), but not boys (n = 48), there was a statistically significant relationship between age at seizure onset and subgroup at 36 months (p < .05), such that girls in the higher scoring group had on average later onset of seizure. Note that although the comparable relationship was in the opposite direction in boys, the interaction between sex and age at seizure onset did not reach significance.

Accumulated seizure burden may more directly capture the impact of epilepsy on neurodevelopment than age at seizure onset. Thus we examined whether cumulative seizure frequency over time was related to neurodevelopmental subgroup at 36 months. Figure 4A shows a heatmap plotting the number of seizure days per week per patient. Those in the lower scoring subgroup were more likely to have more weeks with more seizure days. Focusing on the time interval between the two neurodevelopmental time points examined, patients in the lower scoring subgroup (n = 44) had a median of 152 seizure days from 12 to 36 months (range = 0–646), as compared to median of 0 (range = 0–570) in the higher scoring subgroup (n = 65), a significant difference (Figure 4B, Table 1). Differences were of similar magnitude and significant when accumulated seizure burden either from 0 to 36 months or from 0 to 12 months was used as the independent variable (data not shown).

FIGURE 4.

The lower scoring subgroup had a greater accumulated seizure burden. (A) Heatmap shows individual seizure burden over time measured by the number of seizure days each week. Patients are organized along the y‐axis according to their membership in the subgroups defined by the 36‐month clustergram, such that each row represents one patient. Each square across the x‐axis represents 1 week of study enrollment starting from birth through study completion (0–36 months). Squares are color coded according to the number of seizure days (0–7) for each patient for each week; n = 82 patients have at least one colored square. Blank squares indicate missing data or time outside of the reporting period. Entirely blank rows reflect either patients who reported no seizures (n = 28) or patients who lacked high‐quality seizure diary data (n = 19). (B) Boxplots reflecting the number of accumulated seizure days from 12 to 36 months, stratified by neurodevelopmental subgroup at 36 months. This includes n = 109 patients (all those with high‐quality seizure diary data). (C) Groups further stratified by sex (n = 50 girls and n = 59 boys). *Statistically significant difference.

Because of our post hoc observation of a possible sex difference in the effect of age at seizure, we tested for a sex difference in the effect of accumulated seizure burden. There was a larger difference between the means for the girls in the higher and lower scoring subgroups (173 seizure days) as compared to the comparable difference in the boys (147 days; Figure 4C). However, the interaction (sex × accumulated seizure burden) was not significant by logistic regression (z = −.238, p = .812). The effect of accumulated seizure burden on subgroup was significant separately for both girls and boys (girls: t‐statistic = 3.6492, p = 6.47e‐04; boys: t‐statistic = 3.3619, p = .0014).

Lastly, 36‐month subgroups were analyzed based on neuropsychological test scores at 12 months. There were significant group differences in MSEL (total, and verbal and nonverbal subdomains), VABS‐II, and overall composite 12‐month DQ combining across MSEL and VABS‐II (Table 1).

3.3.1. Logistic regressions

Univariable regressions

All candidate biomarkers of neurodevelopmental subgroup except for 12‐month DQ and ASM exposure were right‐skewed and thus log‐transformed. The single best univariable predictor of neurodevelopmental subgroup was accumulated seizure burden, with 77% classification accuracy (Figure 5A,B). This was more accurate at predicting neurodevelopmental subgroup than 12‐month DQs by 3 percentage points. Age at seizure onset predicted neurodevelopmental subgroup poorly, with accuracy just above chance (54%). Tuber volume achieved a prediction accuracy of 68%. ASM exposure achieved a prediction accuracy of 72%.

FIGURE 5.

Logistic regression results. (A) Table listing the performance characteristics of all logistic regression models, univariable and multivariable. (B) Receiver operating characteristic (ROC) curves for each of the univariable models. (C) ROC curves for the multivariable models. The precision, recall, and F1‐scores are listed separately for the predictions for the higher and lower scoring subgroups. acc, accumulated; ASM, antiseizure medication; AUC, area under the curve; DQ, developmental quotient; sz, seizure.

Model	Accuracy	AUC	Precision		Recall		F1‐score
			High subgroup	Low subgroup	High subgroup	Low subgroup	High subgroup	Low subgroup
Univariable models
acc seizure burden	77%	.80	.80	.72	.82	.70	.81	.71
ASM exposure	72%	.79	.75	.68	.77	.65	.76	.67
12‐month DQ score	73%	.82	.77	.69	.77	.69	.77	.69
Tuber volume	68%	.71	.69	.66	.80	.53	.74	.59
Age at seizure onset	54%	.56	.51	.57	.49	.59	.50	.58
Multivariable models
acc seizure burden, 12‐m DQ score Tuber volume Age at sz onset, ASM exposure	73%	.79	.75	.71	.63	.81	.69	.76
acc seizure burden, 12‐m DQ score Tuber volume Age at sz onset	74%	.79	.77	.72	.63	.84	.70	.77
acc seizure burden, 12‐m DQ score Tuber volume	82%	.86	.86	.76	.83	.80	.84	.78
acc seizure burden, 12‐m DQ score	82%	.86	.85	.77	.85	.77	.85	.77

Multivariable regressions

The five‐variable model had worse performance than the best univariable model, with an accuracy of 73% (Figure 5A,C). Regularization forced the coefficients for age at seizure onset, tuber volume, and ASM exposure to zero in that model, and so we next evaluated a two‐variable solution with accumulated seizure burden and 12‐month DQ as predictors. This model achieved 82% accuracy, slightly outperforming the model with accumulated seizure burden alone. There was no performance boost to the model by adding tuber volume back in.

4. DISCUSSION

This study examined a multicenter cohort of TSC patients longitudinally over the first 3 years of life, focusing on the relationship between neurodevelopment and other clinical features including epilepsy. We observed that children with TSC fall into two large neurodevelopmental subgroups by age 36 months. The children in the lower scoring subgroup (more cognitive and behavioral deficits) were much more likely to have epilepsy and had higher gray matter tuber volumes. Among those with epilepsy, those in the lower scoring subgroup were more likely to have a higher cumulative seizure burden, and among the girls, but not the boys, there was also an increased likelihood of having an earlier age at onset of epilepsy. Notably, neurodevelopmental subgroup at 36 months was, for nearly three quarters of patients, consistent with subgroup at 12 months. In other words, a minority of patients exhibited significant enough differences on neuropsychological testing at age 12 months as compared to 36 months that their subgroup membership flipped over this 2‐year interval. Patients who were consistently in the lower scoring subgroup were more likely to have higher tuber volume, earlier age at seizure onset, lower scores at 12 months, and greater accumulated seizure burden, only the last of which is modifiable between 12 and 36 months of age.

Importantly, we found that accumulated seizure burden was the single best predictor of neurodevelopmental status at 36 months of age in patients with TSC, even more accurate for predicting later development than the same test scores obtained 2 years prior. Combining these two measures (accumulated seizure burden and 12‐month DQ) was more powerful than either variable alone, together predicting 36‐month subgroup with 82% accuracy. Our finding that accumulated seizure burden, rather than age at seizure onset, is most predictive of 36‐month neurodevelopment seems to contradict an earlier report from our group. ⁷ However, methodological differences between the two investigations likely account for the discrepancies between the conclusions. First, Capal et al.⁷ focused on developmental delay at 24 months of age, whereas here we predicted neurodevelopmental subgroup 1 year later, at 36 months. Our analysis also utilized a broader number of inputs (MSEL + VABS‐II vs. MSEL alone) and applied an unsupervised machine learning technique for subgroup definition. Perhaps most significantly, the present analysis utilized a more robust metric of seizure burden. In the Capal study, seizure frequency was defined as a four‐level categorical variable expressing the average number of seizures per month during each 6‐month interval between study visits. Here, we define the continuous variable accumulated seizure burden as the running total number of seizure days (days with at least one seizure). Clearly these two metrics are capturing distinct aspects of the burden of epilepsy. Moreover, the two studies are concordant in de‐emphasizing the importance of infantile spasms as an independent predictor of development in TSC, consistent with results from the PREVeNT trial in which preventative treatment with vigabatrin delayed onset and prevalence of spasms without concomitant improvements in cognitive outcomes at 24 or 36 months. ³¹ Both the present results and the Capal study support the argument that focal seizures, and specifically the accumulation of focal seizures, are detrimental to neurodevelopment in TSC. The current analysis adds that in a head‐to‐head comparison against age at seizure onset, accumulated seizure burden is a more useful proxy for the impact of epilepsy on the developing brain, probably because it captures seizure intractability. Consistent with this premise that uncontrolled focal seizures harm development in TSC, additional recent work from our TACERN study group has shown that the timing of epilepsy surgery matters in young children with TSC. Specifically, earlier surgery (measured in weeks to months) produces greater gains in postoperative language scores. ³² This is in line with an earlier study of TACERN participants that showed that favorable postoperative outcomes in surgical patients were associated with increased MSEL language subscores as compared to medically treated patients. ³³ In the current analysis, a greater proportion of patients in the lower scoring subgroup underwent surgery as compared to the higher scoring subgroup, consistent with their higher cumulative seizure burden. However, the rate of seizure reduction was higher in the higher scoring subgroup (90% vs. 69%), although this difference did not reach significance. Altogether, this underscores how critical it is to consider epilepsy surgery and to consider it early for TSC patients with drug‐refractory epilepsy, with the goal of lowering overall cumulative seizure burden. Not only can postoperative seizure freedom rates be as high as 50%–77% ¹⁰ , ³⁴ , ³⁵ , ³⁶ , ³⁷ with TSC, but the neurodevelopmental benefits of seizure reduction earlier in the infant's disease course may be of even greater consequence for patients with TSC.

Of note, we did evaluate the possibility that the effect of accumulated seizure burden could be confounded by ASM exposure. That is, patients with worse epilepsy are more likely to be taking more ASMs for longer periods of time, and in this cohort we saw that those in the lower scoring subgroup at 36 months had been exposed to a significantly greater number of ASMs than those in the higher scoring subgroup. Nonetheless, ASM exposure was found to be a weaker predictor of cognitive subgroup at 36 months than accumulated seizure burden, performing approximately as well as 12‐month DQ as a single variable. ASM exposure was forced out of the multivariable model by regularization. This suggests that cumulative ASM exposure as a clinical measure does not enhance what is already captured by cumulative seizure burden as a powerful predictor of cognitive otcome in young children with TSC. Although these findings are correlative rather than causative, they are consistent with the argument that trepidations about cognitive side effects of ASMs should be de‐emphasized in this population, and seizure control prioritized.

The post hoc observation that for girls, but not boys, there was an effect of age at seizure onset on neurodevelopment is of great interest. We also observed a qualitatively larger impact of accumulated seizure burden on the girls as compared to the boys. Although we did not anticipate these results, other studies have highlighted sex differences in various TSC manifestations, lending biological plausibility to a possible sex difference in the impact of seizures on development. For example, in one study, male patients were more likely than females to be intellectually disabled and to have renal cysts, retinal lesions, facial angiofibromas, and gingival and ungual fibromas. ³⁸ In other reports, males had more frequent neurologic and eye symptoms and renal cysts, ³⁹ as well as a greater risk of learning disorders and autism. ⁴⁰ In the TSC clinic, pulmonary lymphangioleiomyomatosis is almost entirely restricted to female patients. These sex differences are not completely understood, although it has been shown that sex hormones affect the behavior of cells derived from TSC‐associated lesions. ⁴¹ Larger prospective studies should address the specific possibility of sex differences in the impact of epilepsy on neurodevelopment in TSC. If such effects are replicable, practical clinical consequences would follow.

This study also furnished some noteworthy negative observations. It is well established that pathogenic variants in TSC2 are associated with a greater burden of neurologic symptoms than pathogenic variants in TSC1. ³⁸ , ³⁹ , ⁴² , ⁴³ , ⁴⁴ Accordingly, we showed a trend toward an overrepresentation of TSC1 variants in the higher scoring group. The failure of this effect to reach significance likely reflects a lack of power, because patients with TSC2 variants in our sample outnumbered patients with TSC1 variants more than sixfold. The lack of power in this study highlights a more general practical constraint on relying on genotype for phenotypic prediction in TSC; TSC2 variants are more common in the population and are typically overrepresented in clinical studies. Future models that look in greater detail at the relationship between clinical phenotypes and specific variants in TSC1 and TSC2 are more likely to be useful than models evaluating the impact at the level of which of the two genes is changed.

Additionally, we found that gray matter tuber volume is a reasonable predictor of neurodevelopmental subgroup at 36 months but is not as powerful as accumulated seizure burden, and its inclusion in the multivariable model does not improve the model's prediction. This failure of tuber volume to carry significant prognostic value is consistent with previous studies that have attempted but failed to exploit tuber burden as a reliable clinical predictor. ⁴⁵ , ⁴⁶ , ⁴⁷ , ⁴⁸ , ⁴⁹ , ⁵⁰ , ⁵¹ Currently obtainable structural metrics of tuber burden may simply not be sufficiently precise for useful clinical application.

This study does have limitations. Most notably, we recognize that our readout of seizure accumulation over time relies upon seizure diaries, a choice that necessitates two key caveats. First, there is no incorporation of neurophysiologic (electroencephalographic [EEG]) data into this analysis, so any influence of subclinical seizures or EEG background abnormalities is unrecognized. Second, seizure diaries are subject to problems with reliability. That said, seizure diaries are widely accepted both clinically (as the basis for making adjustments to therapy) and from a research standpoint (as a crucial patient‐reported outcome in both academic and industry‐sponsored studies). Our use of seizure diaries in this study was consistent with the guidelines put forth by the National Institute of Neurological Disorders And Stroke Common Data Elements catalogAlso, we attempted to mitigate the chance of unreliable data by choosing to include only patients with consistently high (>70%) compliance over time, and we would presume that any irregularities or lack of reliability would be randomly distributed within the cohort and thus unlikely to drive the results.

A few less substantial limitations bear mentioning as well. Collection of the data for the natural history patient cohorts from which this study derived began 12 years ago, when treatment practices for children with TSC were slightly different. There is also the possibility of inclusion bias in our study; furthermore, we did not have adequate data to explore the possibility that demographic factors including parental income and educational status could have influenced cognitive outcomes. ⁵² Although our sample size is very good for a rare disease, a larger multisite patient sample would be better to validate the current findings and build upon them by, for example, developing more complex predictive models for neurodevelopment in TSC. Lastly, neuropsychological tests, although well validated, capture only some aspects of how development tangibly impacts functional status. Future predictive models for early life TSC would benefit from incorporation of family‐ and patient‐centered quality of life measures as additional metrics of interest.

5. CONCLUSIONS

This study is the first to disentangle the influence of accumulated seizure burden from that of age at seizure onset on neurodevelopment in very young children with TSC, demonstrating the greater impact of the former than the latter. Importantly, accumulated seizure burden is a metric in these analyses that collapses across all seizure types, including infantile spasms and focal seizures. ASM exposure was not as robust of a predictor of neurodevelopment as accumulated seizure burden. We also showed that neurodevelopmental profiles at age 36 months in TSC fall into two subgroups characterized as higher scoring (less negatively affected) and lower scoring (more negatively affected). Most (~75%) individuals remain in the same subgroup from age 12 to 36 months. Of the factors associated with group switching (tuber volume, age at seizure onset, 12‐month scores, and accumulated seizure burden), only accumulated seizure burden is modifiable. These findings are directly clinically actionable, emphasizing the need to manage all seizures aggressively in the first 3 years of life in TSC, whether medically or surgically.

AUTHOR CONTRIBUTIONS

Conceptualization: S. Katie Z. Ihnen and Samuel Alperin. Data curation: S. Katie Z. Ihnen, Samuel Alperin, Jamie K. Capal, Alexander L. Cohen, Jurriaan M. Peters, and E. Martina Bebin. Formal analysis: S. Katie Z. Ihnen and Samuel Alperin. Funding acquisition: E. Martina Bebin, Hope A. Northrup, Mustafa Sahin, and Darcy A. Krueger. Investigation: S. Katie Z. Ihnen and Samuel Alperin. Methodology: S. Katie Z. Ihnen, Samuel Alperin, Alexander L. Cohen, and Jurriaan M. Peters. Project administration: E. Martina Bebin, Hope A. Northrup, Mustafa Sahin, and Darcy A. Krueger. Resources: E. Martina Bebin, Hope A. Northrup, Mustafa Sahin, and Darcy A. Krueger. Software: Alexander L. Cohen and Jamie K. Capal. Supervision: Darcy A. Krueger. Validation: S. Katie Z. Ihnen. Visualization: S. Katie Z. Ihnen. Writing–original draft: S. Katie Z. Ihnen and Samuel Alperin. Writing–review and editing: Jamie K. Capal, Alexander L. Cohen, Jurriaan M. Peters, E. Martina Bebin, Mustafa Sahin, and Darcy A. Krueger.

FUNDING INFORMATION

Research reported in this publication was supported by the National Institute of Neurological Disorders and Stroke of the National Institutes of Health and Eunice Kennedy Shriver National Institute of Child Health & Human Development under award number U01NS082320. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. M.S. is supported by the Developmental Synaptopathies Consortium (U54NS092090), part of the National Center for Advancing Translational Sciences Rare Diseases Clinical Research Network.

CONFLICT OF INTEREST STATEMENT

J.K.C. serves in the strategic working group for Marinus Pharmaceuticals. M.S. reports grant support from Biogen, Astellas, Bridgebio, and Aucta. He has served on scientific advisory boards for Roche, SpringWorks Therapeutics, Jaguar Therapeutics, Noema, and Alkermes. D.A.K. has received consulting fees from Biocodex, Jazz Pharmaceuticals, and Longboard Pharmaceuticals. He also serves on the board of directors of the TSC Alliance and the medical and scientific advisory committee of the Smith‐Kingsmore Syndrome Foundation. The remaining authors have no conflicts of interest. We confirm that we have read the Journal's position on issues involved in ethical publication and affirm that this report is consistent with those guidelines.

CLINICAL TRIAL REGISTRATION

Clinicaltrials.gov: NCT01780441 and NCT01767779.

Ihnen SKZ, Alperin S, Capal JK, Cohen AL, Peters JM, Bebin EM, et al. Accumulated seizure burden predicts neurodevelopmental outcome at 36 months of age in patients with tuberous sclerosis complex. Epilepsia. 2025;66:117–133. 10.1111/epi.18172

DATA AVAILABILITY STATEMENT

Source data from TACERN and TSC‐EBS are publicly available to external researchers by reasonable request to the TACERN Regulatory Core.