Cognitive Phenotypes in Temporal Lobe Epilepsy utilizing Data- and Clinical-Driven approaches: Moving Towards a New Taxonomy

Abstract

Objective:

To identify cognitive phenotypes in temporal lobe epilepsy (TLE) and test their reproducibility in a large, multisite cohort of patients using both data-driven and clinically-driven approaches.

Method:

Four-hundred and seven patients with TLE who underwent a comprehensive neuropsychological evaluation at one of four epilepsy centers were included. Scores on tests of verbal memory, naming, fluency, executive function, and psychomotor speed were converted into z-scores based on 151 healthy controls (HC). For the data-driven method, cluster analysis (k-means) was used to determine the optimal number of clusters. For the clinically-driven method, impairment was defined as greater than 1.5 standard deviations below the mean of the HC, and patients were classified into groups based on the pattern of impairment.

Results:

Cluster analysis revealed a 3-cluster solution characterized by 1) generalized impairment (29%), 2) language and memory impairment (28%), and no impairment (43%). Based on the clinical criteria, the same broad categories were identified, but with a different distribution; 1) generalized impairment (37%), 2) language and memory impairment (30%), and 3) no impairment (33%). There was a 82.6% concordance rate with good agreement (kappa=.716) between the methods. Forty-eight patients classified as having a normal profile based on cluster analysis, were classified as having generalized impairment (n=16) or an isolated language/memory impairment (n=32) based on the clinical criteria. Patients with generalized impairment had a longer disease duration and patients with no impairment had more years of education. However, patients demonstrating the classic TLE profile (i.e., language & memory impairment) were not more likely to have an earlier age of onset or mesial temporal sclerosis.

Significance:

We validate previous findings from single-site studies that have identified three unique cognitive phenotypes in TLE and offer a means of translating the patterns into a clinical diagnostic criteria, representing a novel taxonomy of neuropsychological status in TLE.

Introduction

Cognitive impairment is the most prevalent comorbidity in patients with temporal lobe epilepsy (TLE), with many patients demonstrating impairments in language, memory, and executive function ^1–3. In a subset of patients, these impairments have been shown to be progressive in nature^4–7. Furthermore, patients with TLE who undergo unilateral anterior temporal lobectomy (ATL) or other surgical procedures are at risk for additional cognitive decline ^{4, 8}. Despite patients with focal TLE having seizures arising from temporal lobe regions, there is variability in the nature and severity of cognitive impairment observed across patients, with some demonstrating generalized impairment, some showing a profile of focal cognitive deficits, and others showing relatively intact cognitive profiles ^9–11.

In efforts to unravel the heterogeneity of cognitive impairment in TLE, studies have shifted from examining TLE patients in aggregate to identifying latent profiles, or cognitive phenotypes, within TLE^9–14. The first study of its kind identified three distinct cognitive phenotypes using cluster analysis, which included a group of patients with isolated memory impairment, a second group with minimal impairment, and a third group with more generalized and pervasive impairment⁹. Follow-up studies have identified similar cognitive phenotypes and shown that these phenotypes are associated with unique patterns of structural and functional abnormalities, with more pervasive cognitive impairment associated with distributed brain abnormalities and isolated deficits associated with restricted regions of brain dysfunction^{9, 10, 11, 12, 14, 15}. However, there is some variability in the phenotypes described across studies, as well as the clinical characteristics and neuroimaging findings associated with each phenotype. This may be due to characteristics of single site data, methods used to derive the phenotypes, the limited sample sizes available, and/or the extent of the cognitive assessment employed (limited versus extensive) and the variability in tests administered across studies. As this literature continues to develop, it is critical to further validate the clinical utility of the cognitive phenotypes derived from data-driven approaches by establishing diagnostic criteria that can be used in clinical practice.

The neuropsychological approach to determining cognitive impairment in clinical practice includes a comprehensive review of all test scores with the operational definition of impairment typically ranging from 1 to 2 standard deviations (SD) below normative means. This approach is employed in presurgical evaluations aimed at estimating risk for postoperative cognitive decline and for determining overall cognitive trajectories in TLE. Recently, our group has utilized a modified clinically-driven method adopted from the mild cognitive impairment (MCI) literature^{10, 13} where phenotypes are derived by considering impairment profiles across multiple tests within each cognitive domain. In this approach, impairment is defined as greater than 1–1.5 SD below the normative mean on two or more measures within each domain and patients are grouped into phenotypes based on the pattern of impairment. Conversely, the most common approach in research studies of phenotyping in TLE has been to derive groups based on cluster analysis^{9, 11, 12, 14}, a data-driven method where objects (e.g., individuals) are portioned into groups based only on information found in the data. The goal of this method is to produce empirically meaningful groups that share common characteristics without restrictions imposed by the user (e.g., clinician).

Given the common use of data-driven approaches in research, we sought to validate the cognitive phenotypes reported in the literature using cluster analysis and then to determine whether the derived data-driven phenotypes are concordant with those identified using a neuropsychological diagnostic approach commonly used in clinical practice. We test both approaches in a large, multi-center cohort of patients with TLE and identify the clinical profiles associated with each phenotype for each approach. Second, we compare the concordance rate between our data-driven and clinical approaches. Based on the existing literature, we predicted that both approaches would yield three phenotypes including a group of patients with generalized impairment, a group with primarily verbal memory and/or language deficits, and third group with normal cognition.

Methods

Participants

This study was approved by the Institutional Review Boards at UC San Diego, UC San Francisco, University of Wisconsin Madison, and Cleveland Clinic. Informed consent was collected from patients and healthy controls (HC) at UC San Diego, UC San Francisco, and University of Wisconsin Madison. At Cleveland Clinic, data were collected as part of an IRB-approved data registry. Four-hundred ninety-four patients with TLE and 150 HC met inclusion criteria for the study. Patients were included in the study if they had a diagnosis of TLE by a board-certified neurologist with expertise in epileptology, in accordance with the criteria defined by the International League Against Epilepsy, and based on video-EEG telemetry, seizure semiology, and/or neuroimaging evaluation. The presence of mesial temporal sclerosis (MTS) was determined by inspection of MRI images by a board-certified neuroradiologist. Healthy controls were recruited through community and patient networks were included if they were between the ages of 18 and 65 and had no reported history of neurological or psychiatric disease.

Neuropsychological measures

All patients and HC completed neuropsychological testing. The following tests were common across the sites and were selected based on recommendations from the National Institute of Neurological Disorders and Stroke (NINDS) Epilepsy Common Data Elements (CDE)¹⁶ and the ILAE Neuropsychology Task Force Diagnostic Methods Commission¹⁷. In addition, measures of motor dexterity and processing speed were included based on previous studies demonstrating that these skills are often impaired in TLE patients with generalized impairment^{9, 11}.Verbal memory was evaluated with Wechsler Memory Scale-Third Edition (WMS-III) Logical Memory (LM) and Verbal Paired Associates (VPA)¹⁸; language ability was evaluated with the Boston Naming Test¹⁹ (BNT) and letter fluency; executive function was measured with the Trail Making Test B (TMT-B); processing speed was measured with TMT-A; fine motor dexterity was measured with Grooved Pegboard [Peg dominant (PegD) and non-dominant hand (PegND)]. Age-corrected scaled scores were calculated for LM and VPA based on normative data provided by the test manual. Age, education, and sex-corrected T-scores were calculated for the BNT, letter fluency, TMT-A, TMT-B, and the Grooved Pegboard based on normative data from expanded Halstead-Reitan Battery²⁰. Although letter fluency has both a language and an executive function component, in our sample letter fluency scores showed a stronger correlation with BNT scores (r = .499, p < .0001) than TMTB (r = .353, p < .001); therefore, we included letter fluency in the language domain. For verbal memory, immediate and delayed memory indices were created by summing the scaled scores for LM I and VPA I immediate total recall scores and LM II and VPA II for delayed total recall, respectively. From a total of 494 TLE patients, 87 patients had missing individual data points on the neuropsychological battery and were therefore excluded from analysis. Case-wise exclusion of patients was necessary given that cluster analysis cannot accommodate missing data. We compared important demographic and clinical variables between the 87 patients that were excluded and the remaining 407 patients. There were no differences in education, age, age of onset, duration of epilepsy, presence of MTS or seizure side (all p-values >.05). Scores for the remaining 407 patients were converted into z-scores based on the HC data. No patients were removed based on outlier detection.

Data-driven: Cluster analysis

Patients’ z-scores were subjected to k-means clustering, an algorithm that defines groups in terms of a centroid, which is typically the mean of a group of points (e.g., cognitive test scores). We tested whether a 3-cluster solution from Hermann et al.⁹ was optimal in our dataset by using the NbClust R package ²¹, which provides 23 indices for determining the number of clusters and proposes the optimal number of clusters from the different results obtained by varying all combinations of number of clusters, distance measures, and clustering methods. To further evaluate the clustering algorithm, the Dunn Index was calculated^{22, 23}. The Dunn Index is the ratio of the smallest distance between observations not in the same cluster to the largest intra-cluster distance. The Dunn Index has a value between zero and infinity, with a higher value indicating better clustering.

Clinically-driven: Neuropsychological criteria

For the clinical approach, we defined impairment as greater than 1.5 standard deviations below the mean of the HC for each test. This impairment cut-off has been shown to be sensitive enough to detect impairment while maintaining specificity²⁴. Second, patients were grouped into phenotypes based on the number of impaired tests and the pattern of impairment, which based on the previous literature was hypothesized to fit three diagnostic cognitive patterns⁵. Generalized impairment was defined as having impairment in at least four of the seven tests, with at least one test per cognitive domain impaired. The domain-specific group was defined as having impairment in verbal memory and/or language, with impairment in either the two tests of verbal memory (i.e., immediate and delayed memory) or the two tests of language (i.e., BNT, letter fluency); for patients with impairment in both domains, they had to be impaired in at least three out of the four tests of verbal memory and language. The normal cognitive profile included patients who met impairment criteria on only one or none of the 7 measures. Several studies have demonstrated that impairment on one test is common among individuals with no neurological or psychiatric disorders ²⁵.

Statistical Analysis

Independent t-tests and Fisher’s Exact tests were used to test for differences in demographic variables between patients and HC. An analysis of agreement using the Cohen’s Kappa statistic was performed to determine the consistency of impairment classification between the two approaches. To determine if the two approaches yield different clinical profiles, analysis of variance (ANOVAs) were conducted to compare clinical and demographic variables across the clinical phenotypes and the clusters, respectively. Benjamini-Hochberg false discovery rate was used to correct for the multiple comparisons across the ANOVAS conducted.

Results

Demographics and patient clinical variables

There were no differences in age [t (555) = .88, p = .38] or sex distribution (Fisher’s Exact = 1.29, p = .289) between patients with TLE and HC; however, as expected, HC had more years of education [t (555) = −6.01, p < .001] (Table 1).

Table 1:

Demographics and clinical variables

	All TLE	HC
N	407	150
Age	36.36 (12.29)	35.31 (13.26)
Education	13.22 (2.4)	14.61 (2.5)
Sex: M/F	182/225	59/91

TLE: temporal lobe epilepsy; F: females; M: males; L: standard deviations are presented inside the parentheses

Cluster analysis

Ten out of the 23 indices from NbClust R package indicated that a 3-cluster solution was an optimal number of clusters for portioning the data. The DI for a 3-cluster solution was DI= 0.098. Cluster 1 was comprised of 29% of patients, Cluster 2 included 28% of the patients, and Cluster 3 was comprised of 43% of the patients (Figure 1). Regarding the pattern of impairment within clusters, patients in Cluster 1 demonstrated impairments across all domains (Cluster-Gen), patients in Cluster 2 showed predominantly impairments in language and/or verbal memory (Cluster-LM), and patients in Cluster 3 demonstrated minimal impairment at the group level (Cluster-MI). Table 2A shows differences in clinical and demographic variables across the clusters. Patients in the Cluster-LM were younger than patients in the Cluster-MI (p<.001) and Cluster-Gen groups (p<.001). Patients in the Cluster-MI had more education relative to the Cluster-LM (p<.001) and the Cluster-Gen (p=.004) groups. Patients in the Cluster-MI also had an older age of seizure onset compared to the Cluster-LM (p <.001), but not with the Cluster-Gen. Patients in the Cluster-Gen had a longer disease duration relative to the Cluster-MI (p=.011) and Cluster-LM (p=.018). The Cluster-MI group had a comparable number of right and left TLE patients, while the Cluster-LM and Cluster-Gen groups both had a greater number left TLE patients.

Figure 1:

Distribution of cognitive scores across groups for the cluster analysis and clinical criteria. Scores are represented as mean z-scores and error bars represent standard deviations.

Table 2A:

Demographics and clinical variables across clusters

	Cluster 1-Gen	Cluster 2-LM	Cluster 3-MI
n	118	113	176	ANOVA	p-value
Age (years)	36.15 (12.44)	34.92 (11.68)	37.43 (12.29)	1.467	.232
Education (Years)	12.97 (2.39)	12.76 (2.19)	13.67 (2.43)	6.012	.003
Age of Onset	15.54 (12.66)	16.89 (12.88)	19.97 (13.91)	4.322	.014
Duration (years)	20.59 (12.82)	18.01 (12.16)	17.43 (12.83)	2.316	.100
				Fisher’s Exact	p-value
Sex: M/F	48/70	60/53	74/102	4.46	.109
Handedness: L/R/A	17/97/3	17/92/4	22/148/6	3.15	.909
Side: L/R/Bilateral	68/35/4	69/32/4	77/67/7	6.89	.135
MTS: Yes/No	53/39	54/41	48/74	9.43	.009

Clinical criteria

Based on the clinical approach, 37% percent of patients met diagnostic criteria for having a generalized impairment profile (Generalized), 30% were impaired on verbal memory and/or language measures (Language & Verbal Memory), and 33% did not meet criteria for impairment and were classified in the no impairment group (No Impairment). No patients showed isolated executive function or processing speed impairments. Out of those patients in the No Impairment group, 52 patients had impairment on one test and 70 patients did not demonstrate impairment on any test. Table 2B shows differences in clinical and demographic variables across the clinically-derived phenotypes. Patients in the No Impairment group had greater years of education relative to the Language & Verbal Memory (p=.001) and Generalized groups (p=.040). Patients in the No Impairment group also had an older age of seizure onset compared to the Language & Verbal Memory (p =.040) and Generalized groups (p= .004). There was a trend for patients in the Generalized group to have a longer disease duration relative to the No Impairment group (p=.050). Information on MTS status was available for 77% of the patient sample. There was a trend for patients in the No Impairment group to have fewer patients with MTS (40%) relative to the other two patient groups.

Table 2B:

Demographics and clinical variables across clinical phenotypes

	No Impairment	Language & Verbal Memory	Generalized
n	133	121	153	ANOVA		p-value
Age (years)	37.53 (12.65)	35.56 (12.17)	35.98 (12.09)	.928	.396
Education (Years)	13.80 (2.43)	12.72 (2.16)	13.10 (2.4)	6.95	.001
Age of Onset	20.99 (13.67)	16.86 (13.61)	15.86 (12.51)	5.81	.003
Duration (years)	16.51 (12.12)	18.69 (12.76)	20.11 (12.94)	2.905	.056
				Fisher’s Exact	p-value
Sex: M/F	61/72	60/61	61/92	2.69	.254
Handedness: L/R/A	20/108/5	15/102/4	21/127/4	2.39	.975
Side: L/R/Bilateral	58/49/3	68/41/4	88/44/8	5.29	.257
MTS: Yes/No	36/54	58/45	61/55	5.53	.063

F: females; M: males; L: left; R: right; A: ambidextrous; MTS: mesial temporal sclerosis; MI: minimal impairment; LM: language & memory; Gen: generalized; standard deviations are presented inside the parentheses

Bold: significant with FDR correction

Concordance

Cohen’s Kappa statistic revealed moderate agreement between the two approaches (κ = .716 p < .001), with a 82.6% concordance rate. Forty-eight patients classified into Cluster-MI (i.e., minimal impairment) met clinical criteria for having verbal memory and language impairment (n=32) or generalized impairment (n=16) (Figure 2). Table 3 shows the clinical characteristics of the clinically-impaired patients that were classified as Cluster-MI with cluster analysis. As expected, these patients demonstrated more subtle or circumscribed impairments across the tests. Specifically, the mis-classified Verbal Memory & Language impaired patients tended to show subthreshold impairment in immediate (z = −1.16) and delayed (z = −1.23) memory rather than in naming (−1.18) or fluency (−.86) (i.e., an isolated verbal memory deficit), and the mis-classified Generalized impairment patients showed impairment in executive functioning (z= −1.85), with more subtle deficits in other domains.

Figure 2.

Distribution of z-scores for patients that were mis-classified as having minimal impairment based on cluster analysis.

Table 3:

Demographics and clinical variables across clinically impaired patients classified as having a normal profile with cluster analysis

	Language & Verbal Memory	Generalized
N	32	16
Age (years)	36.09 (13.19)	37.82 (10.66)
Education (Years)	13.13 (2.34)	14.00 (2.556)
Age of Onset	17.00 (13.82)	16.81 (14.15)
Duration (years)	19.09 (14.27)	21.01 (14.02)
Sex: M/F	13/19	3/13
Handedness: L/R/A	1/30/1	1/14/1
Side: L/R/Bilateral	15/13/2	7/6/2
MTS: Yes/No/Unknown	9/17/6	4/6/6
Neuropsychological Profile
Immediate Memory	−1.16 (.626)	−.74 (.816)
Delayed Memory	−1.23 (.571)	−.771 (.515)
BNT	−1.18 (.1.12)	−1.15 (.928)
Letter Fluency	−.861 (1.09)	−1.02 (1.22)
TMT-A	−.568 (.663)	−1.16 (.867)
TMT-B	−.546 (1.24)	−1.85 (.428)
Peg Dominant	−.714 (.828)	−.865 (.901)
Peg Non-Dominant	−.538 (.898)	−.849 (.777)

F: females; M: males; L: left; R: right; A: ambidextrous; MTS: mesial temporal sclerosis; BNT: Boston Naming Test; TMT-A: Trail Making Test condition A; TMT-B: Trail Making Test condition B; Peg : Grooved Pegboard; standard deviations are presented inside the parentheses

Alternative clusters

In a more recent paper, Elverman et al.¹¹ found a 4-cluster solution to be clinically meaningful, where two groups of patients with focal impairment emerged. Therefore, we also tested the robustness of a 4-cluster solution to examine whether we could identify sub-groups of patients with more focal impairments (see supplemental Figure 1). The 4-cluster solution produced one group that was minimally impaired (26%), one group that showed only language and/or verbal memory impairment (21%), and two groups that showed generalized impairment, but with one disproportionately impaired in language & verbal memory relative to the other (34% and 19%, respectively). However, the latter two groups showed overlapping patterns across most tests. Therefore, the 3-cluster solution produced more distinct, interpretable phenotypes.

Discussion

In a large, multi-site study of 407 patients, we validate previous findings from single-site studies that have identified unique cognitive phenotypes in TLE. We add to this literature by demonstrating the robustness of these phenotypes across data-driven and clinically-driven approaches and, for the first time, show how established neuropsychological criteria can be applied to identify phenotypic impairment at the individual patient level. Previous studies have identified the same general pattern of impairment across phenotypes in TLE: a group with domain-specific impairments in verbal memory and language, a group of patients with broad and pervasive impairment, and a group of patients with intact cognition^9–11. While these studies have been pivotal for discovering these phenotypes in independent datasets, they have not offered a means of translating the patterns into the clinical diagnostic setting. Here, we demonstrate that both approaches identified the same broad phenotypes with moderate agreement. Overall, these findings offer validation that 1) cognitive phenotypes are stable across TLE samples and 2) specific neuropsychological criteria can be applied to re-create these clusters and diagnose impairment profiles at the individual patient level.

Implications of a network disorder in cognitive taxonomy

The traditional view of TLE as characterized by focal, often unilateral, medial temporal lobe dysfunction has been replaced by one that appreciates TLE as a network disorder with widespread abnormalities, including bilateral temporal and extra-temporal cortical thining^26–28, widespread alterations in deep white matter tracks^{29, 30} and superficial white matter^{10, 31, 32}, and increased large-scale network disruption^{33, 34}. More importantly, studies of structure-function relationships in TLE have shown that these diffuse brain abnormalities are associated with a wide range of cognitive deficits^{3, 29, 35}. However, there are inconsistencies across neuroimaging studies in the nature and strength of these associations which may, in part, reflect studies aggregating all patients into one group. We replicate smaller studies that have identified three major cognitive phenotypes within the syndrome of TLE^9–12. Studies examining the neural correlates associated with each phenotype have found that patients with generalized impairment have brain abnormalities that are widespread in nature, those with syndrome-specific memory and language deficits have circumscribed alterations within the temporal lobes, and patients with intact cognition have brains comparable to healthy controls^{10, 14}. Therefore, treating all patients with TLE as a single group may obscure important cognitive and neuroanatomical variability across patient samples that are important for our understanding of the impact of TLE on cognition, which may hamper precision neuropsychology---the diagnosis and treatment of cognitive impairment in patients epilepsy at the individual patient level.

Clinical features associated with cognitive phenotypes

Approximately 25% of patients across recent phenotype studies have demonstrated specific impairments in memory and/or language^9–12. We found that both approaches identified a similar proportion of patients with language and verbal memory impairment, with cluster analysis classifying 28% of patients and the clinical criteria identifying 30% as language/verbal memory impaired. Despite these patients demonstrating the traditional cognitive profile associated with TLE, they comprised the smallest group across both approaches. Furthermore, these patients did not demonstrate the traditional clinical profile associated with TLE; for example, they were not more likely to have an earlier age of seizure onset or a greater proportion of patients with MTS than the other groups. Our generalized group represented 40% of the patients when classified with the clinical criteria. These patients had longer disease duration (20 years on average) relative to patients with isolated verbal memory and language impairment and those with minimally impaired profiles. Longer disease duration has been associated with worse cognition and adverse long-term cognitive outcomes^{4, 5, 36}. Hermann et al.⁹ demonstrated that patients with generalized impairment and longer disease duration were at increased risk for an abnormal cognitive trajectory compared to patients with domain-specific impairment over a 4-year interval. These findings suggest that cognitive phenotyping may not only explain the underlying neurobiology, but may also be important for predicting clinical course. However, disease duration alone may not explain the pervasive cognitive dysfunction observed in these patients given that other studies have not found this association^{10, 11}.Given some evidence that patients with TLE are at increased risk for progressive neurophysiological and structural brain changes, it is possible that patients with generalized impairment represent a group of patients with co-morbid non-epilepsy pathology, elevated health-related risk factors, greater generalized tonic-clinic seizures or low brain reserve. Finally, patients with minimally impaired profiles may represent a group of individuals with higher brain reserve given their intact cognition despite having similar clinical features to those with cognitive dysfunction. In our study, these patients had greater years of education, which has been hypothesized to be protective against epilepsy-related cognitive dysfunction^{37, 38}. Importantly, both approaches produce groups that differed in important clinical and demographic characteristics known to impact cognition.

Misclassification

Although cluster analysis was able to correctly classify 82% of the patients based on their clinically-identified profiles, approximately 12% of patients with clinically significant impairment were classified as having a minimally impaired profile. These patients at the group level had more subthreshold anomalies than those who were classified into the other two groups. These results suggest that cluster analysis may be less sensitive for detecting complex patterns of impairment within smaller samples. It is also possible that given our limited number of tests per domain, we were not able to capture the full pattern of impairment of these patients across all cognitive domains. Notably, cluster analysis is very sample-dependent given that it portions the data based on the information that is available and therefore, it is possible that these patients could have been classified into different solutions if we had a more comprehensive test battery. By contrast, the clinical criteria employed in our study are uniform in nature and are robust to different samples.

Given that Elverman et al.¹¹ identified two subgroups of patients with isolated language and memory impairment and isolated executive function and processing speed, respectively, we ran a 4-cluster solution to determine whether these groups would emerge from those patients that were miss-classified. From our 4-cluster solution, two groups of patients with generalized impairment emerged, with one group demonstrating greater impairment in verbal memory relative to the other group. However, all other tests scores were highly overlapping, limiting the distinctness of these groups. The other two groups identified in the 4-cluster solution were similar to those identified in the 3-cluster solution, which included a group with isolated verbal memory and language impairment and a group with minimal impairment. A potential explanation for the discrepancy between the two studies is that the patients in our study were more impaired, on average, compared to the patients in Elverman et al. In fact, the phenotypes in their study were not clinically impaired but overall demonstrated low scores across multiple domains.

Limitations

There are several limitations to our study. First, we had an appropriate but somewhat limited number of tests per domain and therefore we could not further divide the patients into finer subgroups (i.e., verbal memory only, language only). Second, we did not have common visual memory and visuospatial tests across all three sites and therefore could not include these two domains in our analyses. We recognize that not including a non-verbal memory test in the characterization of TLE phenotypes limits the generalizability of our findings. However, available non-verbal memory tests have shown poor sensitivity to right medial temporal lobe dysfunction in epilepsy ^{1, 39–41}. This poor sensitivity of non-verbal memory measures is reflected in the NINDS Epilepsy CDEs, which do not recommend any specific tests for this domain. Furthermore, the visuospatial domain is included as an optional domain to include and is very seldom impaired in TLE^{39, 41, 42}, even in patients with generalized impairment^{9, 11}. However, future research incorporating protocols that include multiple non-verbal tests that are sensitive to non-dominant hemisphere dysfunction could help to refine cognitive phenotypes in TLE and other epilepsy syndromes. Third, we did not have neuroimaging data on our patients and therefore could not explore brain abnormalities associated with each phenotype. Future studies of brain-behavior relationships with large samples such ours are warranted to replicate the findings in the literature on a large scale. Fourth, we had to remove 87 patients from the analysis given that cluster analysis does not handle missing data. In the future, we plan to explore the misclassification patterns of cluster analysis by comparing this method to other data-driven approaches robust to missing data. Fifth, our patient group consisted of mostly drug-resistant TLE, which may not generalize to all patients with TLE. However, the stability of these three phenotypes has recently been identified in a cohort of patients with mostly well-controlled TLE¹⁵. Finally, there are other epilepsy-related clinical variables (i.e., number and types of seizures, life-time number of GTC seizures, history of AEDs) that were not available in our dataset that may further differentiate the phenotypes identified.

Conclusion

We demonstrate the clinical translation of more than a decade of research into cognitive phenotypes in TLE. Specifically, in a large, multi-center sample, we propose a diagnostic approach for characterizing phenotypic patterns of impairment at the individual patient level. This classification framework not only helps to establish more meaningful cognitive and neurobehavioral taxonomies, but it could improve our ability to predict individual cognitive trajectories and/or match patients to individual treatments in order to improve a range of epilepsy-related outcomes. While studying cognitive dysfunction based on epilepsy syndromes has expanded our understanding of the impact of epilepsy-related pathology on cognition, the phenotyping approach has offered a new classification framework that considers the individual variability observed within and across epilepsy syndromes. Further studies evaluating cognitive phenotypes in other epilepsy syndromes (e.g., frontal lobe epilepsy, genetic generalized epilepsy, juvenile myoclonic epilepsy) are needed in order to identify syndrome-dependent and syndrome-independent phenotypes that could improve our ability to match patients to treatments and improve epilepsy-related outcomes.

Key Points.

• In a large, multi-site study of 407 patients with TLE, we validate smaller single-site studies identifying three unique cognitive phenotypes in TLE.

• We demonstrate that these phenotypes are robust to the methods employed, including clinically-driven and data-driven approaches.

• The data-driven approach misclassified 12% of the patients with clinically-defined significant impairment as having normal cognition.

• Both approaches produce groups that differed in important clinical and demographic characteristics known to impact cognition.

• Cognitive phenotypes offer a new classification framework that considers the individual variability observed within and across epilepsy syndromes.