Lesion and lesion network localization of dysnomia after epilepsy surgery

Abstract

Temporal lobe epilepsy surgery is an effective treatment option for patients with drug-resistant epilepsy. However, neurosurgery poses a risk for cognitive deficits—up to one-third of patients have a decline in naming ability following temporal lobe surgery. In this study, we aimed to better understand the neural correlates associated with reduced naming performance after temporal lobe surgery, with the goal of informing surgical planning strategies to mitigate the risk of dysnomia.

We retrospectively identified 85 patients who underwent temporal lobe resective surgery (49 left temporal lobe, 36 right temporal lobe) for whom naming ability was assessed before and >3 months post-surgery using the Boston Naming Test. We used multivariate lesion-symptom mapping to identify resection sites associated with naming decline, and we used lesion network mapping to evaluate the broader functional and structural connectivity profiles of resection sites associated with naming decline. We validated our findings in an independent cohort of 59 individuals with left temporal lobectomy, along with repeating all analyses after combining the cohorts.

Lesion laterality and location were important predictors of post-surgical naming performance. Naming performance significantly improved after right temporal lobectomy (P = 0.015) while a decrement in performance was observed following left temporal lobectomy (P = 0.002). Declines in naming performance were associated with surgical resection of the left anterior middle temporal gyrus (Brodmann area 21, r = 0.41, P < 0.001), along with a previously implicated basal temporal language area. Resection sites linked to naming decline showed a functional connectivity profile featuring a left-lateralized network closely resembling the extended semantic\default mode network, and a structural connectivity profile featuring major temporo-frontal association white matter tracts coursing through the temporal stem.

This extends prior work by implicating the left anterior middle temporal gyrus in naming decline and provides additional support for the role of the previously identified basal temporal language area in naming decline. Importantly, the structural and functional connectivity profiles of these regions suggest they are key nodes of a broader extended semantic network. Together these regional and network findings may help in surgical planning and discussions of prognosis.

By studying 144 patients who had previously undergone temporal lobe epilepsy surgery, Mhanna et al. found that declines in naming ability post-surgery were associated with removal of key language areas in the left hemisphere. Identifying these areas could aid surgical planning and lead to improved outcomes.

Introduction

Drug-resistant epilepsy constitutes at least 30% of patients with epilepsy.^1,2 Epilepsy surgery can be an effective treatment option for many of these patients.^3,4 The goal of surgery is to achieve better seizure control while minimizing any adverse effects of undergoing surgery. Temporal lobectomy is the most common resective surgery used for the treatment of intractable seizures in adolescents and adults.⁵ It is often effective, with as many as two-thirds of patients being seizure-free more than a year after the surgery.⁶ However, temporal lobectomy bears a potential risk of cognitive deficits, including, but not limited to the domains of memory, visuospatial ability and language. A major focus of current research is understanding the neuroanatomical basis of these post-surgical cognitive changes in hopes of informing improvements in surgical planning that reduce the risk of cognitive impairment following surgery without compromising excellent seizure control.

Language difficulties after left temporal lobectomy can include problems with naming, word finding and verbal fluency. Reduced naming ability is especially common, being seen in approximately one-third of patients after left temporal lobectomy.^7,8 Difficulty naming can be frustrating for patients as it limits one’s ability to communicate clearly in personal and professional settings, impacting patients’ quality of life and potentially reducing interest in social engagement.⁹ However, the predictors of naming decline after temporal lobectomy and the temporal lobe regions implicated in naming are still unclear. While candidate anatomical sites have been identified, they are not always consistent across studies.

The left temporal lobe includes key nodes of distributed brain networks involved in naming.^10,11 In recent decades, an anatomical framework has emerged that highlights ventral and anterior temporal cortex as supporting multi-modal semantic and conceptual associations that are essential for normal naming performance. Impaired word retrieval has been linked to lesions of the left anterior and posterior lateral temporal cortex in neurological patients.^10,12,13 Using perfusion imaging methods, DeLeon et al.¹⁴ showed that numerous left hemisphere (LH) areas were crucial for object naming in a group of patients with acute ischaemic stroke. Specifically, naming impairment was associated with dysfunction of the left anterior temporal lobe, including the superior and middle temporal gyri.¹⁴ Fonseca et al.¹⁵ further showed that decreased metabolism in posterior basal and lateral temporal areas was correlated with poorer naming in temporal lobe epilepsy. There is, therefore, strong reason to believe that anterior temporal lobe structures play an important role in the neural processes underlying naming.

Lesion-symptom mapping is particularly relevant to understanding which brain regions, when surgically resected, are associated with long-term deficits in naming. This approach is clinically important, as many regions with functional MRI activity patterns that strongly correlate with naming performance may not be necessary for naming, while a subset of regions will be critical for naming and lesions to those regions will be associated with long-term deficits. Three recent studies have shown that resection of a basal temporal language area, located mainly in the fusiform gyrus and surrounding parahippocampal gyrus, is significantly associated with naming decline following left temporal lobe epilepsy surgery.^11,16,17 Interestingly, these temporal lobectomy lesion studies have not implicated lateral temporal cortex, which is also thought to be a critical region for naming.^18,19 For example, Baldo and colleagues²⁰ showed that the left mid-posterior middle temporal gyrus and underlying white matter play a critical role in object or picture word retrieval in a cohort of patients with left hemisphere stroke. Similarly, Bowren and colleagues¹² implicated left anterolateral temporal cortex, among other distributed left hemisphere regions, as important for naming in a cohort of 432 individuals with focal brain lesions.

Across different lesion aetiologies, naming decline has been associated with lesions of both ventral and lateral temporal regions. One possibility is that these lesion sites are intersecting with distinct nodes of a common network. One way to investigate this possibility is with lesion network mapping, which infers the anatomically distributed brain networks that are interrupted by a lesion. Lesion network mapping leverages connectivity information from healthy individuals to infer the structural and functional connectivity network effects of focal lesions.^21,22 To our knowledge, this approach has not yet been applied to naming decline in association with temporal lobe epilepsy surgery.

In the current study, we aimed to identify the predictors of naming performance after temporal lobe epilepsy surgery in 85 drug-resistant epilepsy patients. Naming was evaluated using the Boston Naming Test (BNT) before surgery and in the chronic epoch after the surgery (>3 months). We hypothesized that there will be a decline in naming after left temporal lobe surgery and that this would be correlated with the resection of critical regions of the left anterior temporal lobe.^10,12,13 We did not pre-specify an anatomical region of interest beyond this level of resolution given that both ventral and lateral temporal regions have been implicated in naming. Instead, we applied data-driven multivariate lesion-symptom mapping to identify areas of the left temporal lobe that, when resected, are significantly associated with a decline in naming performance. In addition, we hypothesized that the critical regions for naming within the left temporal lobe would be connected with a distributed brain-wide network that supports naming, which we investigated with lesion network mapping. To evaluate the generalizability of our findings, we evaluated whether models trained on lesion-symptom and lesion network maps obtained from the primary patient cohort from the University of Iowa (i.e. Iowa cohort) could be used to predict naming outcomes in an independent epilepsy temporal lobectomy cohort from the multisite FATES study.¹¹ We also evaluated predictions in the reverse direction—training models on the lesion-symptom and lesion network maps from the FATES cohort and using them to predict naming outcomes in the Iowa cohort. Lastly, we investigated whether other lesion and patient characteristics are associated with postoperative naming performance decline.

Materials and methods

Participants

The primary study sample included 85 participants with drug-resistant epilepsy who underwent temporal lobe resections (49 left temporal lobe and 36 right temporal lobe). Participants were identified from the Patient Registry of the Division of Behavioral Neurology and Cognitive Neuroscience within the Department of Neurology, as well as the Department of Neurosurgery at the University of Iowa. The cohort was selected from a pool of 135 patients with drug refractory epilepsy who underwent temporal lobectomy between 1992 and 2021. Patients were included in the analysis if they completed both pre- and post-surgical BNT assessments in the chronic epoch (>3 months) and had post-surgery brain imaging (MRI or CT scan). Using these criteria, 50 patients were excluded. Forty-nine patients did not have both pre- and post-surgery BNT scores and one patient did not have post-surgery brain imaging. Our institutional ethical review committee approved the study design in advance of data collection and analysis. The demographic characteristics of the sample are presented in Supplementary Table 1.

Ethics approval

All procedures performed in the study were in accordance with the University of Iowa Institutional Review Board.

Consent to participate

Informed consents were obtained from all participants in accordance with the University of Iowa Institutional Review Board.

Naming assessment

Word retrieval was assessed using the standard 60-item version of the BNT before surgery and at least 3 months following surgery. The BNT is a measure of confrontational word retrieval, classically used to evaluate patients with aphasia.²³ It consists of 60 line drawings of objects, graded in difficulty. An item is considered passed if the patient either spontaneously names the item correctly within 20 s or names the item correctly within 20 s after being given a stimulus (semantic) cue from the examiner. The assessments were performed with standard testing and scoring procedures. The BNT change score was computed by subtracting the preoperative from the postoperative score. We also calculated the per cent change score using the following equation: [100 × (postoperative score − preoperative score) / preoperative score].

Brain imaging and lesion segmentation

Neuroimaging analyses were performed based on the postoperative high-resolution volumetric T₁-weighted MRI (n = 83) or CT (n = 2) scan. All lesions were manually traced upon reviewing the lesion location in three dimensions and the anatomical accuracy of the lesion boundaries was reviewed and edited as needed in native and MNI space by a neurologist blinded to naming outcomes, as previously described.²⁴ Lesion tracing encompassed all tissue that was resected. Any perilesional or remote collateral damage was included in the lesion mask. This includes peri-lesional gliosis or peri-lesional haemorrhage if the damage was evident on follow-up structural imaging. For patients who registered in the Patient Registry prior to 2006, the surgical resection cavity was manually drawn for each participant using the MAP-3 lesion tracing method, which entails manually tracing lesion borders on a template brain.^25,26 Following 2006, lesions were manually drawn on native T₁-weighted or CT images²⁷ and then converted to the 1-mm MNI152 template brain using a high deformation, non-linear, enantiomorphic, registration procedure from the Advanced Normalization Tools.^28-30

Lesion-symptom mapping

We applied multivariate lesion-symptom mapping to the behavioural and lesion data across all 85 patients to identify brain structures that, when resected, are significantly associated with absolute changes in BNT. We used the LESYMAP package in R (https://github.com/dorianps/LESYMAP), which uses sparse canonical correlation analysis for neuroimaging (SCCAN).³¹ The SCCAN method uses an optimization procedure to assign a continuous weight to voxels (range 0–1) that maximize the multivariate correlation between voxel weights and behavioural scores while imposing a sparsity constraint on the final solution. The optimal sparseness and validity of the map is evaluated using a repeated 4-fold, within-sample cross-validation; a model is built using 75% of the sample and then applied to the remaining 25% to predict the BNT change scores from lesion location. The correlation between model predicted and actual BNT change scores is used to evaluate the statistical significance of the overall map thresholded at the optimal sparseness. Because significance testing occurs at the level of the entire map, this procedure avoids the need for voxel-wise multiple comparisons. Coordinates of significant clusters were displayed on the MNI152 template brain using Slicer software for 3D visualization.

Lesion network mapping

We evaluated the connectivity patterns of resection sites associated with post-resection naming decline using lesion network mapping. This approach infers brain-wide connectivity patterns of lesion sites using ‘connectome’ data from healthy individuals.²¹ We performed these analyses for both (i) the grey matter, using resting state functional connectivity MRI (rs-fcMRI); and (ii) the white matter, using diffusion MRI tractography.

Functional lesion network mapping

Functional lesion network mapping was performed to evaluate which brain regions exhibit greater connectivity to resection sites associated with naming decline. We performed this analysis in two ways. First, each lesion mask was used as a single seed in a rs-fcMRI analysis. The resulting 85 lesion-derived networks were then evaluated in relation to naming performance using a voxel-wise permutation analysis of linear models (FSL PALM; https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/Randomise). This approach identifies statistical relationships between voxel-wise functional connectivity values and BNT change, as previously performed.^32,33 Two-tailed significance tests were performed using cluster-free threshold enhancement (10 000 permutations) with family-wise error correction.

In addition to using voxel-wise linear models, we also seeded networks from peak regions identified from the lesion-symptom map result, as performed previously.^12,33,34 This differs from using each individual lesion mask to ‘seed’ a network analysis, and avoids some of the problems associated with signal averaging from large lesions.²¹ For example, larger lesions may include signals from many regions not associated with the outcomes of interest, leading to signal mixing that can ‘wash out’ signal from relevant regions. To avoid this problem, the lesion-symptom map can be used to select anatomical regions that are most likely to be associated with the specific symptom of interest, thus increasing the likelihood of identifying a network causally related to the behaviour of interest. To identify peak regions of interest (ROIs) with the strongest association with post-resection naming decline, we identified the peak grey matter region within the lesion-symptom map generated from our analysis. A 1-cm ROI in this region was used to seed a functional connectivity analysis. For both approaches, the same rs-fcMRI data were used from the Genome Superstruct Project (n = 1000), as in previous work.^33,35 The rs-fcMRI data were processed in accordance with previously described methods and included global signal regression.³³ For each of the 1000 concatenated fcMRI files, the temporal correlation between the average resting state blood oxygen level-dependent (BOLD) signal within the seed ROI and the BOLD signal of every other voxel within the brain was calculated to identify regions with positive and negative correlations. Pearson correlation coefficients were transformed to Fisher’s z-coefficients using the variance-stabilizing Fisher r-to-z transformation. Each of the resulting 1000 individual z-score maps were then combined into a single group-derived t-map. For the lesion-seeded analysis, this yielded a single group-derived t-map for each patient’s lesion, which was then entered in the FSL PALM analysis. For the peak-seeded analysis, this yielded a single group-derived t-map reflecting the connectivity profile of the peak ROI from the lesion-symptom map.

Structural lesion network mapping

To evaluate structural networks associated with lesion location, each lesion mask was used to ‘seed’ a deterministic tractography analysis using Lead-DBS software, as performed previously.³³ Streamline connectivity profiles associated with each lesion were estimated using a normative dataset of neurologically healthy individuals with high quality diffusion MRI data from the Human Connectome Project’s MGH 32-fold group connectome, as used in previous work.^36,37 For structural lesion network mapping, the challenges of signal averaging within a large lesion mask that occur with rs-fcMRI do not apply, and thus only the lesion-seeded approach was used. The direction of streamlines was not constrained to any other ROI beyond the starting ‘seed’ ROI. The unthresholded lesion-derived streamline tractography maps generated from each individual patient’s lesion were then evaluated using FSL PALM, as described in the previous section for the lesion-seeded functional lesion network mapping analysis.

Out-of-sample predictions

Our study design included an independent cohort of 59 patients enrolled in the Functional MRI in Anterior Temporal Epilepsy Surgery (FATES) study¹¹ who underwent resections in the left temporal lobe for drug-resistant focal epilepsy and who completed the BNT before and 7 months after surgery. Our intention with this cohort was to evaluate whether the lesion-symptom mapping and lesion network mapping results from the Iowa cohort could be used to predict the relative severity of naming impairment or lack of impairment in out-of-sample naming performance, i.e. to validate the findings from the Iowa cohort in an independent patient sample. To aid in the interpretation of the lesion-derived models, we first trained a model on demographic and non-lesion location variables that were available in both cohorts. This included age, education, lesion volume, presurgical BNT scores and sex. We created a model from the Iowa cohort data and then used it to predict change in BNT scores in FATES. Next, the relationship between lesion location and naming performance was modelled using a series of linear regressions generated using the Iowa cohort data. Lesion loads were defined for each map generated in the Iowa cohort (i.e. lesion-symptom map, functional lesion network map, structural lesion network map) by summing the voxel intensities across all voxels that intersected with each patient’s lesion, and these lesion loads were used as predictors to build linear regression models linking lesion location to BNT change scores. Lesion loads were computed in the same way for the FATES cohort (using the maps derived from the Iowa cohort) and the regression models trained in the Iowa cohort were then applied to the lesion load predictors defined for the FATES cohort to obtain predicted BNT change scores. Out-of-sample prediction performance was quantified by calculating the Pearson correlation of predicted and observed BNT change scores in the FATES cohort; if the model generalizes, then the predicted naming change scores should strongly correlate with the observed naming change scores. We note that this is not a measure of overall predictive performance as it is not sensitive to constant additive offsets or differences in scaling between the predicted and observed values, but only reflects the extent to which the predictions are a linear function of the true behavioural scores. A lesion volume model was also fit to the Iowa cohort and applied to the FATES cohort, allowing for the comparison of the lesion load models to a model that lacked information about lesion location and included only information about lesion size. In addition to using the results from the Iowa cohort to predict naming outcomes in the FATES cohort, we also performed the same analyses in reverse—i.e. we conducted the same lesion-symptom mapping and lesion network analyses in the FATES cohort and then used these results to build models that were used to predict naming outcomes in the Iowa cohort.

Combined Iowa and FATES analyses

A final use of the FATES data was to conduct a combined analysis using data from all subjects in both cohorts to maximize the sample size and statistical power of the analyses.

Relationship to previous studies

As noted earlier, prior lesion studies of temporal lobectomy-associated dysnomia have implicated different regions. We evaluated whether peak regional findings from previously published lesion-symptom mapping studies of naming outcomes were associated with worse naming outcomes in the Iowa cohort (Table 1).^11,16,17 We produced 1-cm spherical ROIs at the peak sites associated with naming decline in the lesion-symptom mapping analysis, divided our cohort into those with lesion masks that intersected or did not intersect with these ROIs, and compared the post-surgical naming outcome between the two groups using independent samples t-tests. One-way ANCOVAs were conducted to determine if there was a statistically significant difference between the intersecting and non-intersecting groups on the naming change while adjusting for potential differences in lesion volume.

Table 1.

Temporal lobectomy studies using lesion-symptom mapping to identify neural correlates of dysnomia

Study	Number of participants	Diagnosis	Lesion-symptom map peak (MNI coordinates) in the MNI 152 standard-space
Snyder et al.,17	95	Left temporal lobe epilepsy surgery	(−31.64, −12.34, −32.13)
Reindl et al.,16	108	Epilepsy surgery	ITG peak (−52, −12, −37), FG peak (−38, −21, −32)
Binder et al.,11	59	Left temporal lobe epilepsy surgery	(−40.79, −34.15, −26.28)

We also used the peak coordinates of our study combined with other published lesion-symptom mapping studies^11,16,17 to seed rs-fcMRI networks. Our goal here was to evaluate if lesion studies implicating different temporal lobe regions may map to a common brain network. To evaluate this question, we generated network maps from the peak coordinate or coordinates in each previous study (using a 1-cm spherical ROI as a seed) and used principal component analysis (PCA) to decompose the resulting brain-wide network maps into latent variance components. This was done to evaluate how much of the variance in network topography across the different studies could be explained by a single, principal network versus multiple, distinct networks.

Evaluation of non-lesion variables in association with naming decline

In addition to lesion-location analyses, we also performed stepwise multiple linear regression analyses to identify patient and other lesion characteristics that are associated with pre- to post-surgery changes in naming ability in the right and left hemisphere surgery groups, analysed separately. We included 11 independent variables in the models: age at surgery, age at epilepsy onset, pre-surgery BNT score, time of naming assessment after surgery, seizure freedom after surgery based on Engel classification, lesion volume and years of education. We also included baseline cognitive performance in other non-naming domains: Controlled Oral Word Association Test (COWA), Benton Facial Recognition Test (BFRT), Wide Range Achievement Test (WRAT-Reading Test) and Wechsler Adult Intelligence Scale (WAIS) Similarities Test. Variables were retained in the model if the P-value ≤ 0.05. Statistics of demographic and behavioural data were performed using IBM SPSS Statistics 20.0 (http://www.spss.com).

Results

Behavioural performance findings

Patients with left temporal lobe surgery had a significant decline in naming performance after surgery, with 57% showing lower postoperative score compared to baseline. They had an average decline of 3.43 points [standard deviation (SD) = 7.4, P = 0.002; or 6.04%, SD = 17%]. Thirty-seven per cent had a decline of 10% or more in their naming score, 14% had a decline of 20% or more and six patients (12%) showed a decline of >30%. In contrast, patients with right temporal lobe surgery on average had an improvement in the BNT score postoperatively. A majority of patients (64%) had a higher postoperative score with an average increase of 1.7 points (SD = 3.9, P = 0.015; or 4.26%, SD = 10.56%) with a broad range (−17% to 50%). Results for each patient are shown in Supplementary Table 2.

Lesion-symptom mapping

Lesion overlap across all patients is shown in Fig. 1A. The lesion-symptom mapping analysis revealed that BNT decline was most strongly associated with lesions of the left anterior middle temporal gyrus (Brodmann area 21), with a peak voxel at MNI coordinate (−59, 3, −27) (overall map r = 0.44, P ≤ 0.001; Fig. 1B). Those individuals in the Iowa cohort with lesions that intersected with this coordinate had a 7-point greater reduction in naming performance relative to subjects with non-intersecting lesions (n = 42 and 43, respectively; see Supplementary Table 3 for additional details). There did not appear to be any specific category of items that uniquely contributed to this regional peak; patients with lesions intersecting the peak finding had significant reductions in naming performance in each of the BNT subcategories when analysed individually (animals, plants and tools; P < 0.001, 0.002 and <0.001, respectively). While naming performance was reduced in each subcategory, we noted tools to decline significantly more than other subcategories (Supplementary Table 4). Additional clusters were observed in the lesion-symptom map, including the white matter of the temporal stem, at the confluence of the inferior fronto-occipital fasciculus and the uncinate fasciculus (MNI −36, −7, −18), the anteromedial temporal cortex partially overlapping the amygdala (MNI −29, 0, −22) and in the basal temporal language area on the anterior part of the fusiform gyrus near the occipitotemporal sulcus (MNI −39, −14, −24). To ensure these regional findings were not unique to the multivariate lesion-symptom mapping method used, we repeated the analyses using a mass univariate analysis in which the Pearson point-biserial correlation was computed between lesion status and BNT change scores separately for each voxel with permutation-based significance testing (10 000 permutation iterations) continuous family-wise error (FWE) rate control, which showed the same topography of regional findings that reached statistical significance (Supplementary Fig. 1).

Figure 1.

Lesion overlap and lesion-symptom mapping results. (A) The lesion overlap map of the surgical resection cavities used in this analysis. Maximum overlap is 47, located in the anterior temporal lobe. The colour scale range is 0–47. (B) The multivariate lesion-symptom mapping results of the Iowa cohort (n = 85, r = 0.41, P < 0.001). The strongest association was between post-surgical naming decline and resection of the left anterior middle temporal gyrus. The colour scale range is 0–1. (C and D) The multivariate lesion-symptom mapping results of the combined (Iowa and FATES) left temporal lobectomy cohorts (n = 108, r = 0.55, P ≤ 0.001). The strongest association was between Boston Naming Test (BNT) decline and resection of the left anterior fusiform gyrus, extending laterally to the inferior temporal and middle temporal gyri, and medially to the parahippocampal gyrus. The colour scale range is 0–1. L = left; R = right.

Lesion network mapping

Functional lesion network mapping

The functional connectivity networks derived using either of the two complementary functional lesion network mapping methods were similar. They highlighted a left-lateralized network spanning the basal temporal language area, the anterolateral middle temporal gyrus, and white matter extending from these temporal regions to the temporal stem white matter. The results obtained using FSL PALM indicated that inferred connectivity from lesion locations to the regions shown in Fig. 2A and B was associated with a reduction in naming performance (FWE-corrected P < 0.05). This can be interpreted as showing regions that preferentially ‘disconnect’ with the resection cavity when naming performance declines, though with the important caveat that we did not measure evidence of network disconnections in the patients themselves, and this is based on inference from normative connectivity information. The peak site of lesion connectivity associated with reduction in naming performance was associated with the white matter temporal stem at the confluence of the inferior longitudinal fasciculus (ILF), and the inferior fronto-occipital fasciculus (IFOF) (MNI −32, −8, −10), extending into the anterior fusiform temporal cortex and to the inferior and middle temporal gyri laterally, and the parahippocampal gyrus medially. The unthresholded brain-wide network is shown in Fig. 2C. The network derived from the peak voxel in the Iowa cohort lesion-symptom map (seed ROI from result displayed in Fig. 1B surface view) is shown in Fig. 2D. Overall, the broader network topography resembled what has been described as the extended semantic network (displayed in Fig. 2E for reference)³⁸ and the default mode network B in the Yeo 17 atlas.³⁹

Figure 2.

Functional lesion network mapping results. (A and B) The significant functional network connectivity results of the Iowa cohort projected onto axial and surface views, respectively. Each lesion mask was used as a seed in this analysis, with results shown reaching statistical significance at family-wise error (few)-corrected P < 0.05. The strongest finding was at the white matter temporal stem (MNI −32, −8, −10), extending into the anterior fusiform temporal cortex, and to the inferior and middle temporal gyri laterally, and the parahippocampal gyrus medially. The colour scale range is 1482.54 to 2008.78. (C) The same network map as in A and B without a statistical threshold applied. Network results associated with greater dysnomia are shown in red. (D) A network map derived from the peak site in the lesion-symptom map shown above in Fig. 1B surface view. Notably, the core network closely resembles what has been described as the extended semantic network.³⁸ (E) The extended semantic network displayed for reference. (F) A network derived from the Functional MRI in Anterior Temporal Epilepsy Surgery (FATES) peak lesion-symptom map site. While the peak region was in a different region of the temporal lobe, the network pattern resembles that of the Iowa sample derived using the same approach (seen in D). The colour scale range is −23.20 to 23.20. (G) A network map derived from our results as well as peak regions of lesion-symptom maps from the literature, decomposed into orthogonal variance components using principal component analysis with the observed principal component explaining 64% of the overall variance. The colour scale range is −8.63 to 8.63. (H) A lesion network map of Boston Naming Test performance derived previously by our group from an independent sample of 432 neurological patients with focal, acquired lesions.¹² This network also demonstrates the core nodes of the extended semantic network. The colour scale range is 0–18.13.

Structural lesion network mapping

Figure 3A shows the significant white matter findings of the structural lesion network map, representing streamlines with significantly greater connectivity to lesions associated with a decline in naming performance (FWE-corrected P < 0.05). Spatial correlations between the map shown in Fig. 3A and the common white matter tracts from the HCP-842 tractography atlas⁴⁰ were calculated. Significant white matter findings correlated most strongly with four major white matter association tracts in the temporal lobe, including the left ILF (Fig. 3B), the IFOF (Fig. 3C), the arcuate fasciculus (AF) (Fig. 3D) and the uncinate fasciculus (UF) (Fig. 3E).

Figure 3.

Structural lesion network mapping results and how they align with white matter tracts. (A) Significant white matter findings from the voxel-wise permutational analysis of linear models (PALM) associated with post-surgery naming decline in the Iowa cohort. The colour scale range is 0–2391. (B) The overlap between our findings and the left inferior longitudinal fasciculus (ILF) (outlined in green). (C) The overlap between our findings and the left inferior fronto-occipital fasciculus (IFOF) (outlined in green). (D) The overlap between our findings and the left arcuate fasciculus (AF) (outlined in green). The masks of the ILF, IFOF and AF white matter tracts are derived from HCP-842 tractography atlas.⁴⁰ (E) The overlap between our findings and the left uncinate fasciculus (UF) (outlined in green). (F) White matter regions with significantly higher connectivity to lesion sites with reduced naming performance in the Functional MRI in Anterior Temporal Epilepsy Surgery (FATES) cohort. The colour scale range is 0–3110.

Out-of-sample prediction

We evaluated whether these results generated using the Iowa cohort predict individual differences in naming performance in an independent cohort (FATES). First, using the demographic and non-lesion location variables, we show the predicted and observed FATES BNT change scores were not significantly correlated from models derived using Iowa data (r = 0.13, P = 0.34). Next, we show that predictions from models fit on lesion loads for each lesion-derived map does significantly correlate with observed changes in naming performance when applied to the FATES cohort (P < 0.001; Table 2).

Table 2.

Lesion derived maps of BNT change predicting BNT change scores in validation cohorts

	Iowa predicting FATES		FATES predicting Iowa
	r	P	r	P
Non-lesion models	0.13	0.34	0.03	0.77
LSM	0.51	<0.001*	0.44	<0.001*
sLNM	0.61	<0.001*	0.32	0.003*
fLNM	0.59	<0.001*	0.23	0.035*
Lesion volume	−0.63	<0.001*	−0.05	0.66

r = correlation between the predicted and actual Boston Naming Test (BNT) change scores; FATES = Functional MRI in Anterior Temporal Epilepsy Surgery; LSM = lesion-symptom mapping; sLNM = structural lesion network mapping; fLNM = functional lesion network mapping.

^*Statistical significance.

We performed the lesion-symptom mapping and lesion network mapping independently using the FATES data. This resulted in a significant lesion-symptom map (r = 0.66, P < 0.001) with a peak association in the anterior ventral temporal lobe, centred on the fusiform gyrus (MNI −40, −23, −21), which is located very close (∼1 cm) to the original results reported using a different lesion-symptom mapping approach.¹¹ A functional lesion network map using PALM showed significant voxels in the precentral and fusiform gyri, with peak MNI coordinates of −13, −18, 63 and −43, −27, −20, respectively (Supplementary Fig. 2). A brain-wide network derived from the peak of the lesion-symptom map was partially consistent with the network derived from Iowa cohort in that it highlighted key nodes of the extended semantic network, including the inferior parietal lobe and anterior temporal cortex (Fig. 2F). The structural lesion network map of the FATES cohort was similar to that of Iowa cohort and intersected with each of the four major white matter tracts, ILF, IFOF, AF and UF (Fig. 3F). Using these results from FATES, we also evaluated predictions in the Iowa cohort. We first evaluated a model based on demographic and non-lesion variables from FATES to predict Iowa BNT change scores in the Iowa cohort and showed a non-significant correlation of predicted and observed scores (r = 0.03, P = 0.77), indicating that these variables together did not predict BNT change scores. Models trained on the lesion-derived maps were used to generate predicted naming change scores that did significantly correlate with the observed change scores. Models trained on lesion volume alone did not show a significant correlation of actual and observed BNT change scores (Table 2).

Combined Iowa and FATES analyses

We also performed lesion symptom mapping and lesion network mapping on combined data from the Iowa and FATES cohorts. For lesion-symptom mapping, the full map was again statistically significant (r = 0.55, P ≤ 0.001), and both the basal temporal language area and the anterolateral portion of the middle temporal gyrus were represented, which corresponded to the regional peaks obtained from FATES and Iowa when run independently. The strongest association was between BNT decline and resection of the left anterior fusiform gyrus (−39, −27, −21), extending laterally to the inferior temporal and middle temporal gyri, and medially to the parahippocampal gyrus (Fig. 1C–E). The cluster on the anterolateral middle temporal gyrus had a peak MNI coordinate of (−61, −2, −27). The PALM-derived functional lesion network map included regions represented in the Iowa and FATES maps, with regions associated with reduced naming performance showing significantly greater connectivity with the anterior fusiform gyrus (MNI −42, −28, −20), extending laterally to the inferior and middle temporal gyri, and medially to the parahippocampal gyrus (all voxels displayed are FWE-corrected P < 0.05, Supplementary Fig. 3A and B). The overall brain topography is best seen with the unthresholded functional connectivity map, which closely resembled the result derived from Iowa (spatial correlation r = 0.85, Supplementary Fig. 3C). The structural lesion network maps from the combined cohort appeared very similar with those of the Iowa and FATES cohort separately (Supplementary Fig. 3D).

Relationship to prior studies

Effects of lesion intersection with peak sites previously associated with naming decline by independent studies (Table 1) were also evaluated. We compared naming outcomes between subjects with resections that overlapped with these peak regions and those that did not. Individuals with lesions intersecting with these ‘dysnomia risk’ sites uniformly showed greater naming decline than those who did not, in some cases reaching statistical significance (Fig. 4, actual values and comparison of results using ROIs of varying size shown in Supplementary Table 5 and Supplementary Fig. 4). When controlling for lesion volume, significant results hold. Additionally, patients with lesions intersecting Binder et al.¹¹ peak showed a significant reduction in the BNT change compared with those with non-intersecting lesions [F(1,46) = 6.16, P = 0.01].

Figure 4.

Evaluation of lesion-symptom mapping results from prior studies. Naming outcome after left temporal lobe (TL) surgery was evaluated in our cohort in relation to lesion intersection with other peak sites associated with dysnomia from previous studies. Individuals with lesions intersecting with these peak lesion-symptom mapping results uniformly showed greater naming decline after surgery than those who did not, in some cases reaching statistical significance. MNI coordinates for regional peaks include: Snyder et al.¹⁷ (MNI −31.64, −12.34, −32.13), Reindl et al.¹⁶ ITG (inferior temporal gyrus) peak (MNI −52, −12, −37), Reindl et al.¹⁶ FG (fusiform gyrus) peak (MNI −38, −21, −32) and Binder et al.¹¹ (MNI −40.79, −34.15, −26.28). The error bars represent the standard error of the mean (SEM). BNT = Boston Naming Test.

We also generated functional connectivity network maps seeded by each of these peak lesion-symptom mapping coordinates from our study and these other published lesion-symptom mapping studies (Supplementary Fig. 5). We used PCA of these brain-wide network maps to generate a composite network and evaluate how much of the variation across maps can be explained by a single network map, as performed previously.^12,33 Figure 2G shows the first principal component, which captured 64% of the variance across maps and looked quite similar to the networks derived from Iowa and other studies independently. Finally, we show that this same extended semantic network was identified in a previous functional lesion network mapping study by our group¹² that used a non-overlapping sample of 432 participants with focal, acquired brain lesions that were not limited to the temporal lobe, demonstrating that this network result is also robust to non-epilepsy and non-temporal lobe lesions (Fig. 2H).

Evaluation of non-lesion variables in association with naming decline

For the LH group, patients with greater postoperative decline in naming performance were significantly older at the time of epilepsy onset (B = −0.15, P < 0.03), and this variable explained 9% of the variance in the naming outcomes (adjusted R² = 0.09). For the right temporal lobe group, there was a significant association between the naming outcome and baseline naming performance, the time of naming assessment post-surgery, and the WRAT reading performance. This model explained 65% of the variance in naming score change (adjusted R² = 0.65). Right temporal lobe patients with higher baseline performance showed a greater decline in naming postoperatively (B = −0.43, P < 0.001). For each point increase in preoperative naming score, there was a corresponding 0.43-point decline in the naming score following right temporal lobe surgery. Patients with later naming assessments after surgery also demonstrated greater improvement in naming performance (B = 0.52, P = 0.01) and right temporal lobe patients with better performance in WRAT reading showed greater improvement in postoperative naming scores (B = 0.14, P = 0.001). There was no significant association between a change in naming performance and any of the following factors for right or left temporal lobe surgery groups, analysed independently: age at surgery, seizure freedom after surgery, lesion volume, years of education, and baseline cognitive performance in non-naming domains (COWA, BFRT and WAIS similarities).

Discussion

This study investigated the factors associated with naming performance after temporal lobe surgery in patients with drug-resistant epilepsy. Overall, our results build upon earlier findings showing postoperative naming decline associated with left temporal lobe surgery.^{11,15-17,41-44} These previous studies have highlighted critical anatomical nodes for naming that have prominently included the fusiform gyrus, which has been referred to as the basal temporal language area.^45-48 Our peak regional lesion-symptom map result for the combined analysis (Iowa + FATES) was within this basal temporal language area, specifically at the anterior fusiform gyrus and extending to the adjacent parahippocampal gyrus (n_combined = 108, r = 0.55, P ≤ 0.001). Our results also extend the existing literature by highlighting an additional region of the left anterior middle temporal gyrus (Brodmann area 21) that, when resected, was also significantly associated with a decline in naming performance. The anterolateral middle temporal gyrus lesion-symptom map from the Iowa cohort (Fig. 1B) has not been reported in prior temporal lobectomy lesion-symptom mapping studies, to our knowledge. However, this region has been implicated in the semantic and lexical processing involved in naming in prior lesion and functional imaging work.^{10,12-14,49-51} In fact, the significant finding displayed in Fig. 1B overlaps with a prior lesion-symptom map of naming performance in an independent (non-overlapping) cohort of 432 individuals published by our group.¹² These different regions of the left temporal lobe language networks (anterior middle temporal gyrus versus fusiform gyrus) may contribute differentially to naming, but both appear to be involved in semantic processing and have connectivity with the extended semantic network.^{13,15,20,51-60}

Our functional lesion network mapping results suggest that this region is part of a larger network associated with naming, overlapping with what has previously been called the extended semantic network or default mode network B.^38,39 Lesions that disrupt this network are associated with dysnomia. This is demonstrated across independent analyses derived from multiple temporal epilepsy cohorts. In fact, it appears this extended semantic network may tie together previously disparate anatomical findings from different lesion-symptom mapping studies, as the network can be derived from regional peaks in the basal temporal language area and from the anterolateral middle temporal gyrus. While individual studies have identified discrete regions associated with dysnomia, our network derived from individual coordinates spanning ventral and lateral temporal regions indicate these regional results are likely part of a single broader network important for naming. As such, the network map appears to support a more unified view of lesions involving the extended semantic network being associated with chronic dysnomia after temporal lobectomy. Further, this finding does not appear to be limited to epilepsy patients or temporal lobe lesions, as the same network was demonstrated in an independent, non-temporal lobectomy cohort (Fig. 2H).¹²

The extended semantic network was named based on its purported role in semantic function; it includes anatomical nodes in the inferior parietal lobe, superior frontal gyrus, and antero-ventral temporal lobe.^38,61-63 This semantic network is connected by different major association white matter tracts, including the ILF, AF and IFOF.⁶⁴ Taken together, the current findings and those from the literature provide supporting data for a prominent role of the grey and white matter of a semantic network that supports naming, and specifically, the semantic cognitive processes involved in the complex task of naming.^14,65

While anatomy and connectivity of the resection cavity is an important factor in naming outcomes, it likely interacts with other factors. We observed that in left hemisphere temporal lobectomy, later onset of epilepsy was associated with a greater reduction in naming ability relative to individuals with earlier onset of epilepsy. This is in line with previous research.^66-73 Stafiniak and colleagues⁷⁴ found that age of onset of epilepsy after age 5 was associated with a significant decline (≥25%) in naming after left temporal lobe resection compared to subjects with an earlier age of onset of epilepsy (≤5 years). In addition, Reindl and colleagues¹⁶ showed decline in naming performance 6 months after dominant temporal lobe surgery in patients with epilepsy onset after 5 years of age but not with earlier epilepsy onset. This may be attributed to brain plasticity in patients who experience epileptic seizures at an early age,⁷⁵ which could facilitate language reorganization and thus greater resilience to surgical resection of the seizure focus.^66,69,70,76

Another finding we wish to highlight is the improvement in naming performance observed after right temporal lobe epilepsy surgery. To the best of our knowledge, this was the first study that showed significant improvement in naming performance after right temporal lobectomy, possibly because naming is less commonly studied in association with right hemisphere surgery. Possible explanations for this include fewer epileptiform discharges, less frequent seizures, fewer side effects of seizure medications or, most likely, some combination of these factors. A few studies investigated the effect of right temporal lobe resection on visual naming, with findings indicating no significant change in BNT performance^73,77 or in the naming of living and non-living entities.⁷⁸

Our study has limitations. The study included a wide range of patients in terms of onset of epilepsy, age at the time of surgery, and there was variable timing after surgery for assessing naming, ranging from 3 months up to 9 years. We focused on the BNT and did not include other categories of visual naming (e.g. famous faces, landmarks) or other modalities (e.g. auditory naming). While BNT is a validated tool to assess naming, it is not sensitive to naming deficits that can be overcome within 20 s, even though this latency may be quite disruptive to the normal use of language. Further, we used correlation of observed and predicted naming to assess out-of-sample model performance, but it may be possible in future studies to include cohorts with comparable changes in naming performance and assess the predictive performance of numeric changes in naming performance. In addition, future studies could evaluate the utility of network disconnection information acquired directly from the patients, which may explain additional variance in outcomes beyond what is possible using normative data.^79,80 In addition, our study did not distinguish whether the decline in naming ability after left temporal lobe surgery is due to impairments in concept identification (semantic) or word retrieval (lexical), or a combination of the two.

In summary, our study highlighted the role of lesion location in postoperative naming outcomes following temporal lobe epilepsy surgery. Our results strongly implicate the basal temporal language area in naming. In addition, we identified the left anterior middle temporal gyrus as a critical region for naming in patients with drug-resistant epilepsy. These findings were present in a discovery dataset with supporting evidence from an independent validation cohort. Together, these regions appear to both be part of an extended brain-wide semantic network associated with naming. Given that significant ‘out-of-sample’ variance in naming performance can be explained using lesion-symptom and lesion network maps, it may be possible to continue developing these maps using larger cohorts, such that they eventually can be utilized for pre-surgical planning. Given the results to date, it may be possible to transform these lesion-symptom maps onto pre-surgical anatomical scans to help avoid resection of the anterior temporal middle gyrus and basal temporal language area whenever possible. Further, precision mapping of the temporal lobe semantic network nodes in individual patients may aid in the planning of pre-surgical language mapping. Moreover, it may be possible to map a planned surgical resection cavity with reference to these maps to assess the risk of naming deficits.

Data availability

The datasets generated and/or analysed during the current study are available upon request from the corresponding author.

Funding

This study was supported by the National Institute of Neurological Disease and Stroke (1 R01 NS114405-02; 5R01DC004290-22). This work was conducted on an MRI instrument funded by 1S10OD025025-01.

Competing interests

The authors report no competing interests.

Supplementary material

Supplementary material is available at Brain online.