Task-based functional magnetic resonance imaging prediction of postsurgical cognitive outcomes in temporal lobe epilepsy: A systematic review, meta-analysis, and new data

Abstract

Task-based functional magnetic resonance imaging (tfMRI) has developed as a common alternative in epilepsy surgery to the intracarotid amobarbital procedure, also known as the Wada procedure. Prior studies have implicated tfMRI as a comparable predictor of postsurgical cognitive outcomes. However, the predictive validity of tfMRI has not been established. This preregistered systematic review and meta-analysis (CRD42020183563) synthesizes the literature predicting postsurgical cognitive outcomes in temporal lobe epilepsy (TLE) using tfMRI. The PubMed and PsycINFO literature databases were queried for English-language articles published between January 1, 2009 and December 31, 2020 associating tfMRI laterality indices or symmetry of task activation with outcomes in TLE. Their references were reviewed for additional relevant literature, and unpublished data from our center were incorporated. Nineteen studies were included in the meta-analysis. tfMRI studies predicted postsurgical cognitive outcomes in left TLE ($\widehat{\rho }=-.27$, 95% confidence interval [CI] = −.32 to −.23) but not right TLE ($\widehat{\rho }=-.02$, 95% CI = −.08 to .03). Among studies of left TLE, language tfMRI studies were more robustly predictive of postsurgical cognitive outcomes ($\widehat{\rho }=-.27$, 95% CI = −.33 to −.20) than memory tfMRI studies ($\widehat{\rho }=-.27$, 95% CI = −.43 to −.11). Further moderation by cognitive outcome domain indicated language tfMRI predicted confrontation naming ($\widehat{\rho }=-.32$, 95% CI = −.41 to −.22) and verbal memory ($\widehat{\rho }=-.26$, 95% CI = −.35 to −.17) outcomes, whereas memory tfMRI forecasted only verbal memory outcomes ($\widehat{\rho }=-.37$, 95% CI = −.57 to −.18). Surgery type, birth sex, level of education, age at onset, disease duration, and hemispheric language dominance moderated study outcomes. Sensitivity analyses suggested the interval of postsurgical follow-up, and reporting and methodological practices influenced study outcomes as well. These findings intimate tfMRI is a modest predictor of outcomes in left TLE that should be considered in the context of a larger surgical workup.

Task-based functional magnetic resonance imaging (tfMRI) has developed as a common alternative in epilepsy surgery to the intracarotid amobarbital procedure (IAP), also known as the Wada procedure.^1,2 Multiple prior research syntheses have demonstrated diagnostic concordance between tfMRI and IAP,^2–6 implicating tfMRI as a comparable predictor of postsurgical cognitive outcomes. A review of the current literature conducted by the American Academy of Neurology Guideline Development, Dissemination, and Implementation Subcommittee assessed issues regarding the added value of tfMRI for the prediction of cognitive outcomes after anterior temporal lobectomy (ATL).² They concluded that language fMRI has achieved Level C evidence (“possible” benefit), whereas the prognostication of memory status by memory fMRI has reached Level B evidence (“probable” benefit). Although qualitative, systematic reviews have been published,^3,4 attempts to quantitively assess the current literature have not been successful.⁷

The results of research syntheses may be attributed to methodological heterogeneity and reporting biases within the literature. Because the wider tfMRI literature has been historically populated with studies involving small sample sizes and low power,⁸ it is conceivable that studies vary considerably on statistical grounds as well.⁹ Taken together, studies substantiating the predictive validity of tfMRI have been inconclusive in the setting of epilepsy surgery. Moreover, no quantitative estimates summarizing its predictive validity have been introduced. Thus, we sought to systematically review and quantitatively synthesize the literature predicting postsurgical cognitive outcomes in patients with temporal lobe epilepsy (TLE) using tfMRI. To bolster meta-analytic sample size, we also introduce new data from our center.

1 |. MATERIALS AND METHODS

1.1 |. New data

1.1.1 |. Participants

Our retrospective analysis included 156 patients with medically refractory unilateral TLE referred from the Thomas Jefferson University Comprehensive Epilepsy Center and scanned between 2008 and 2020 as part of a presurgical evaluation for temporal lobe resection or laser interstitial thermal therapy. A temporal lobe seizure focus and surgical candidacy were determined by review of medical history and examination, scalp video encephalography, MRI, positron emission tomography, and neuropsychological assessment in all patients. In addition to clinical confirmation of TLE, study eligibility involved having completed a verb generation fMRI task on a single scanner and neuropsychological evaluation (NPE) within 12 months before surgery and between 6 and 36 months following surgery. Of the 156 patients referred, a total of 53 were included in the final sample (Figure S1).

1.1.2 |. Measures

NPEs were performed on a clinical basis before and after surgery to evaluate changes in confrontation naming (Boston Naming Test), verbal fluency (Controlled Oral Word Association Test), verbal memory (California Verbal Learning Test–Second Edition Long Delayed Free Recall), and nonverbal memory (Wechsler Memory Scales–Third Edition Faces II). Seizure freedom, defined as attaining an Engel Class I or not, was evaluated with a modified Engel surgical outcome scale.¹⁰ Additional information regarding the follow-up conducted is in the Supplementary Material (p. 1; Figure S1).

1.1.3 |. MRI data acquisition and data processing

All patients underwent a presurgical fMRI scan including a verb generation task on a 3.0-T Philips Achieva scanner with an eight-channel head coil. Additional information about the structural and functional sequence parameters and verb generation tfMRI paradigm are available in the Supplementary Material (p. 1–2). Imaging data was preprocessed using fMRIPrep (version 20.2.1)¹¹ based on Nypipe (version 1.5.1).¹² Subsequent image smoothing and postprocessing was done in Statistical Parametric Mapping (SPM) 12 (revision 7771),¹³ and bootstrapped laterality indices (LIs) were obtained with the LI toolbox extension.¹⁴ Additional information regarding image processing can be found in the Supplementary Material (p. 1–5).

1.2 |. Systematic review and meta-analysis

1.2.1 |. Protocol and registration

This review was reported according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines.¹⁵ The protocol associated with this review is available at the PROSPERO systematic review registry (CRD42020183563; https://www.crd.york.ac.uk/prospero/display_record.php?ID=CRD42020183563).¹⁶

1.2.2 |. Eligibility criteria

This review was limited to studies predicting postsurgical cognitive outcomes using tfMRI in patients with TLE. Reports involving patients with extratemporal, multifocal, or generalized epilepsies were excluded, although TLE surgical approaches were inclusively operationalized. Both prospective and retrospective studies were considered eligible. To maintain internal validity and comparison with prior studies of the IAP, studies solely examining unilateral regions of interest (ROIs) were excluded. Thus, only studies utilizing an LI or similar approach were incorporated. Only studies reporting sufficient data were included in the meta-analysis, namely effect sizes and sample sizes at minimum. Signed p-values were used to calculate effect sizes if other effect sizes were either unreported or unusable.

1.2.3 |. Information sources and literature search

The PubMed and PsycINFO databases were queried for English-language articles published between January 1, 2009 and December 31, 2020. The last search was on April 8, 2021. Citations within identified primary and secondary literature were also used to find relevant work. Data from our cohort reported here were also included. Search terms applied to both database searches are available in the Supplementary Material (p. 6).

1.2.4 |. Study selection

The literature searches were conducted by the first author and two research assistants (A.B.-S., D.W.). The relevance of studies to the present review were evaluated following examination of article title, abstract, and methods by the first author and two research assistants (A.T., Y.R.). The first author and senior author determined final eligibility for inclusion.

1.2.5 |. Data collection process and data items

The first author and year were obtained, each article was assigned a numeric identifier, and each individual effect size reported within eligible studies was assigned an additional identifier. Effect sizes, test statistics, and p-values and their corresponding sample sizes were retrieved from eligible studies at minimum. Follow-up interval, ROI, hemispheric language dominance, intellectual functioning, postsurgical cognitive decline, surgery type, seizure freedom rating, age at epilepsy onset, and epilepsy duration were recorded where provided. Additional variables were recorded, including country of origin, age, sex, education, stereotactic space, scanner used, field strength, and analytic software package. Effect sizes were entered along with coding of fMRI task domain and cognitive assessment domain. Finally, effect sizes were coded such that negative effects corresponded with poorer cognitive outcomes following epilepsy surgery among patients with normative left hemisphere language dominance. Efforts were made to contact authors (n = 8) of studies reporting data insufficient to be included in the meta-analysis; however, four authors were unreachable, and the remaining authors were unable to provide the appropriate data from their original studies.

1.2.6 |. Risk of bias assessment

A tailored Newcastle–Ottawa Scale (NOS)¹⁷ for cohort studies was completed for each study by the first and senior authors. Because eligible studies were to include patients with TLE and did not require a comparison group, items referring to the selection of a nonexposed group (Selection category, Item #2), demonstrations that outcomes were not present prior to enrollment (Selection category, Item #4), and comparability of cohorts (Comparability category, Item #1) were skipped. Our resulting scale had a valid score range of 0 to 5.

1.2.7 |. Statistical analysis

The statistical methods used to analyze our sample are available in the Supplementary Material (p. 5–6). Interrater agreement of NOS ratings was determined with Cohen kappa¹⁸ and Kendall rank correlation.¹⁹ Before formal meta-analysis, effect sizes were first converted to Pearson correlations where necessary, then standardized using a Fisher z-transformation.²⁰ Spearman correlations were converted to Pearson correlations following Pearson.²¹ Other effect sizes were converted according to Friedman.²² A random-effects meta-analysis with a restricted maximum likelihood estimator of heterogeneity was conducted.²³ Subsequently, moderation and meta-regression (linear mixed-effects) analyses were performed with clinical and sociodemographic factors previously described. The alpha level was set to .05 for all analyses, and post hoc pairwise contrasts were Holm-adjusted where appropriate. All analyses were performed in the R language (version 4.2.0)²⁴ with the metafor package (version 3.4–0).²⁵

1.2.8 |. Publication bias

Conventional graphical and quantitative tests were used to evaluate publication bias. Cochran Q-test²⁶ and estimates of between-study variance (I2)²⁷ were used to examine heterogeneity. Funnel plots, Begg and Mazumdar rank correlation,²⁸ Egger and colleagues’ linear regression,²⁹ and Orwin fail-safe N for effect size³⁰ were used to determine publication bias. Egger regression test was performed with studies’ standard errors as the predictor, and Orwin fail-safe N was estimated based on a target effect size of −.05, an average effect of limited practical significance. Duval and Tweedie’s trim-and-fill method³¹ was used to adjust model estimates for publication bias. Sensitivity analyses of study reporting and methodological variables were also conducted with linear mixed-effects approaches.

1.2.9 |. Certainty of evidence

The certainty of evidence was based on consensus between the first and senior authors and appraised according to the Grading of Recommendations, Assessment, Development, and Evaluations (GRADE) framework.³²

1.2.10 |. Institutional review board statement

Our primary study was approved by the Thomas Jefferson University institutional review board, and all participants provided written informed consent.

2 |. RESULTS

2.1 |. New data

Demographic and correlational statistics from our sample are presented in the Supplementary Material (Table S1).

2.2 |. Study selection

Thirty-eight articles were reviewed for formal eligibility, and 18 articles were included in the meta-analysis (Figure 1).^33–50 Included articles were first published between 2003 and 2020 (Table 1), yielding a total of 146 effects. The inclusion of our unpublished data increased this effect size count to 154 for a total of 19 studies included. Excluded references and a complete PRISMA checklist are available in the Appendices.

FIGURE 1.

PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) flowchart

Table 1.

Characteristics of Included Studies

Study	Country	Sample	n	Surgery Type	Patient Selection	Follow-up Interval (Mo.)	MR Field Strength	tfMRI Task	ROI	tfMRI Independent Measure	Outcome Measure	Total NOS Rating a
Sabsevitz et al. (2003)	United States	LTLE	24	ATL	Consecutive	6.00	1.5T	Semantic Tone Decision	Temporal, Frontal, Angular Gyrus, Hemisphere	LI	Boston Naming Test	5
		RTLE	32
Rabin et al. (2004) b	United States	LTLE	10	ATL	Consecutive	6.90	1.5T	Scene Encoding and Recognition	Hippocampus, Mesial Temporal	LI	Scene Memory Test	4
		RTLE	13
Janzsky et al. (2005) c	Germany, Hungary, Switzerland	RTLE	16	ATL	Consecutive	6.00	1.5T	Roland Hometown Walking	Hippocampus	LI	Rey Visual Design Learning Test	4
Powell et al. (2008)	United Kingdom	RTLE	7	ATL	Consecutive	3.00	1.5T	Memory Encoding	Hippocampus	Activation Symmetry (Difference Image)	Verbal Learning, Design Learning	3.5
Frings et al. (2008)	Germany	LTLE	12	Mixed (ATL, LINAC, VNS)	Consecutive	3.00	1.5T	Spatial Memory	Hippocampus	LI	Verbaler Lern & Merkfaehigkeitstest, Diagnosticum für Cerebralschädigung	3.5
		RTLE	10	ATL
Binder et al. (2008)	United States	LTLE	60	ATL	Consecutive	6.00	1.5T, 3.0T	Semantic Tone Decision	Hemisphere, Frontal, Temporal	LI	Selective Reminding Test, Wechsler Memory Scale-Revised/Wechsler Memory Scale III Logical Memory	4
		RTLE	62
Binder et al. (2010) d	United States	LTLE	30	ATL	Consecutive	6.00	1.5T	Semantic Tone Decision	Hippocampus	LI	Selective Reminding Test, Wechsler Memory Scale-Revised/Wechsler Memory Scale III Logical Memory	4
		RTLE	37
Bonelli et al. (2010)	United Kingdom	LTLE	41	ATL	Unspecified	4.00	3.0T	Material Encoding, Verbal Fluency, Verb Generation, Reading Comprehension	Hippocampus, Hemisphere	LI, Activation Symmetry (Asymmetry Image)	Verbal Learning, Design Learning	4
		RTLE	31
Dupont et al. (2010) b	France	LTLE	11	ATL	Consecutive	6.00	1.5T	Memory Encoding and Recognition	Hippocampus	LI	Jones-Gotman Verbal Learning Test, Rey Auditory Verbal Learning Test, Rey-Taylor Complex Figure Test, Aggie Figure Learning Test	5
		RTLE	13
Bonelli et al. (2012)	United Kingdom	LTLE	24	ATL	Unspecified	4.00	3.0T	Verbal Fluency	Inferior Frontal Gyrus, Medial Temporal Gyrus, Hippocampus	LI	McKenna Graded Naming Test	3.5
		RTLE	20
Janecek et al. (2013) b	United States	LTLE	10	ATL	Consecutive	6.00	1.5T, 3.0T	Semantic Tone Decision	Hemisphere	LI	Boston Naming Test	4
Rosazza et al., (2013) e	Italy	LTLE	42	ATL	Retrospective	12.00	1.5T	Naming, Verb Generation, Verbal Fluency	Inferior Frontal Gyrus, Temporal	LI	Boston Naming Test, Letter Fluency, Semantic Fluency	5
Sidhu et al. (2015) b	United Kingdom	LTLE	23	ATL	Prospective Cohort	4.00	3.0T	Letter Fluency, Verb Generation, Reading Comprehension, Memory Encoding	Frontal, Frontotemporal	LI	BIRT Memory and Information Processing Battery List Learning	4
		RTLE	27
Strandberg et al. (2017) b	Sweden	LTLE	8	ATL	Retrospective	13.56	3.0T	Verbal Memory Encoding, Roland Hometown Walking	Inferior Frontal Gyrus, Medial Temporal Gyrus	LI	Claeson-Dahl Learning and Retention Test, Rey-Osterrieth Complex Figure Test	5
		RTLE	6			9.33
Audrian et al. (2018)	Canada	LTLE	20	ATL	Mixed (Prospective Cohort, Retrospective)	12.00	3.0T	Verb Generation, Sentence Completion, Category Fluency, Naming to Description	Hemisphere, Inferior Frontal Gyrus, Temporal	LI	Boston Naming Test	4
You et al. (2019) b	United States	LTLE	20	Mixed (ATL, SAH)	Retrospective	15.60	3.0T	Auditory Description Decision	Inferior Frontal Gyrus, Posterior Temporal Gyrus	LI	Boston Naming Test	4
		RTLE	15	Mixed (ATL, SAH, Lesionectomy)
Trimmel et al. (2019)	United Kingdom	LTLE	25	Mixed (ATL, Lesionectomy)	Consecutive	4.00	3.0T	Auditory Naming, Picture Naming, Verbal Fluency	Temporal, Frontal	LI	McKenna Graded Naming Test	4
		RTLE	21
Foesleitner et al. (2020)	Austria	LTLE	12	Mixed (ATL, SAH, Transsylvian SAH, Lesionectomy)	Retrospective	17.00	3.0T	Verb Generation, Semantic Tone Decision	Inferior Frontal Gyrus, Superior Temporal Gyrus, Hemisphere	LI	Boston Naming Test, Regensburg Verbal Fluency Test	3.5
		RTLE	16
Present Study	United States	LTLE	35	Mixed (ATL, Modified ATL, Posterior ATL, LITT)	Retrospective	15.26	3.0T	Verb Generation	Hemisphere	LI	Boston Naming Test, Controlled Oral Word Association Test, California Verbal Learning Test-Second Edition, Wechsler Memory Scale-III Faces	5
		RTLE	18	Mixed (ATL, LITT)		15.27

Note. tfMRI = Task-based functional magnetic resonance imaging. MR = Magnetic resonance. ROI = Region of interest. LTLE = Left temporal lobe epilepsy. RTLE = Right temporal lobe epilepsy. ATL = Anterior temporal lobectomy. LINAC = Medical linear acceleration. VNS = Vagal nerve stimulation. SAH = Selective amygdalohippocampectomy. LITT = Laser interstitial thermal therapy. LI = Laterality Index.

^aTotal NOS rating reflects the interrater average.

^bReanalyzed from published data.

^cThe non-dominant hemisphere resection group was included only.

^dBinder et al. (2010) reported on a subsample of Binder et al. (2008). Thus, only a subset of data reported were included.

^eTwelve-month outcome data included only.

2.3 |. Study characteristics

From the 19 studies included in the meta-analysis, 17 independent samples of left TLE (total n = 407, median across studies = 23, range = 8–6 0) and 16 independent samples of right TLE (total n = 344, median across studies = 17, range = 6–62) patients were drawn. Within these samples, patient groups received primarily ATL (n = 24), with the remaining receiving a mixture of surgery interventions (n = 9). The average follow-up interval was 7.80 months (median = 6, range = 3–17). Additional summary characteristics of included patient samples are in Table 2.

Table 2.

Patient Characteristics across Included Studies

	Mean	Median	Range
Age (Years)a	38.41	38.00	34.63 – 48.88
Sex (% Male)	45.02%	43.75%	15.38% – 66.67%
Education (Years)	13.72	13.40	11.6 – 15.6
Handedness (% Right)	84.24%	86.67%	40.00% – 100.00%
Age of Epilepsy Onset (Years)	16.73	16.00	9.55 – 30.17
Duration of Epilepsy (Years)	21.43	21.20	11.00 – 32.55
Seizure Freedom (% Seizure Free)	71.57%	71.11%	43.75% – 100.00%
General IQ	94.49	93.61	91.9 – 99.71
Performance IQ	95.54	95.78	93.14 – 98.53
Verbal IQ	94.66	93.50	88.43 – 110.77
Language Dominance (% Left Hemisphere)	89.15%	91.67%	65.00% – 100.00%
Postsurgical Cognitive Decline (% Declined)	31.61%	30.00%	0.00% – 63.64%

^aReported age at enrollment, scan, or surgery.

SPM was mostly frequently employed to analyze tfMRI data (n = 24), followed by Analysis of Functional NeuroImages (AFNI; n = 7) and Voxbo (n = 2). Studies also varied regarding the method of stereotactic normalization during image preprocessing, particularly, studies were normalized to a Montreal Neurological Institute (MNI) template (n = 12), individual patient brains (i.e., no normalization was used, n = 12), a scanner-specific template (n = 4), a group average template (n = 3), or the Talairach template (n = 2).

Studies reported an average of 4.67 effect sizes (median = 3, range = 1–15). Among effect sizes reported, they were primarily Pearson correlations (k = 76), followed by null effects (k = 58), Fisher Z-statistics (k = 8), Spearman rank correlations (k = 7), and p-values (k = 5).

2.4 |. Risk of bias assessment

Average total NOS ratings across studies was 4.11 (average difference between raters = .11, SE = .13). Interrater agreement between raters was substantial (average k = .71, SE = .19) and ratings were highly correlated (Τ = .69, p = .001). Item-level ratings from each rater are available in the Supplementary Data.

2.5 |. Meta-Analysis

Heterogeneous results were observed among tfMRI studies predicting postsurgical cognitive outcomes in TLE (Q[153] = 252.05, p < .0001). Among studies, between-study variance was low (I² = 37.06%). Nonetheless, effect sizes followed an approximate normal distribution (Figure S2). Meta-analysis demonstrated a small overall effect across studies ($\widehat{\rho }=-.16$, SE = .02, Z = −7.01, p < .0001, 95% confidence interval [CI] = −.21 to −.12).

2.6 |. Sensitivity analyses

2.6.1 |. Follow-up interval

A linear mixed-effects analysis demonstrated a significant association between follow-up interval and effect size outcomes (k = 154, Q_M[1] = 10.79, Z = 3.28, p = .001), which moderately accounted for study outcomes (pseudo R² = 9.52%). However, significant heterogeneity remained (Q_E[82] = 239.10, p < .0001), suggesting additional factors are accounting for effect size variances.

2.6.2 |. Year of publication

A trend-level association between year of publication and effect size outcomes was observed, but this association was not significant (k = 146, Q_M[1] = 3.22, p = .07). Note our unpublished data were not included in these analyses, as it was not applicable.

2.6.3 |. Reported effect size type

With Pearson correlations set as the reference level, effect size outcomes varied significantly among reported effect size types (k = 154, Q_M[4] = 102.45, p < .0001), and no significant residual heterogeneity remained (Q_E[149] =149.60, p = .47). Post hoc contrasts suggested all factor levels differed significantly from one another (Q_M[2] ≥ 29.11, p ≤ .0002). Studies reporting Spearman rank correlations (k = 7, $\widehat{\rho }=-.43$, SE = .10, Z = −4.38, p < .0001, 95% CI = −.62 to −.24) were most negative in direction, followed by studies reporting Fisher Z statistics (k = 8, $\widehat{\rho }=-.27$, SE = .08, Z = −3.50, p = .0005, 95% CI = −.42 to −.12) and Pearson correlations (k = 76, $\widehat{\rho }=-.24$, SE = .03, Z = −9.76, p < .0001, 95% CI = −.29 to −.19). Studies reporting signed p-values and null effects were, on average, in the opposite direction from studies reporting other effect size types; studies reporting signed p-values were more positive in magnitude (k = 5, $\widehat{\rho }=.36$, SE = .18, Z = −1.96, p = .051, 95% CI = −.00 to .71) than studies reporting null effects (k = 58, $\widehat{\rho }=.24$, SE = .04, Z = 6.71, p < .0001, 95% CI = .17 to .31).

2.6.4 |. Magnetic field strength

Magnetic field strength within studies moderated effect size outcomes (k = 154, Q_M[3] =50.10, p < .0001). Post hoc contrasts indicated each group of studies varied when compared to one another (Q_M[2] ≥ 29.04, p < .0001). Studies involving 1.5-T scanners demonstrated the most negative outcomes (k = 37, $\widehat{\rho }=-.24$, SE = .05, Z = −4.59, p < .0001, 95% CI = −.34 to −.14), followed by studies corporating both 1.5-T and 3.0-T scanners (k = 31, $\widehat{\rho }=-.15$, SE = .04, Z = −3.89, p < .0001, 95% CI = −.23 to −.08) versus only 3.0-T scanners (k = 86, $\widehat{\rho }=-.13$, SE = .04, Z = −3.73, p = .0002, 95% CI = −.20 to −.06). This analysis also showed significant residual heterogeneity remained (Q_E[151] = 250.31, p < .0001).

2.6.5 |. Analysis software

Effect size outcomes were moderated by fMRI image analysis software choices (k = 154, Q_M[3] = 50.20, p < .0001). Post hoc comparisons indicated each group of studies varied when compared to one another (Q_M[2] ≥ 20.76, p < .0001). Studies utilizing SPM resulted in the most negative outcomes (k = 105, $\widehat{\rho }=-.18$, SE = .03, Z = −5.43, p < .0001, 95% CI = −.24 to −.11), followed by studies using AFNI (k = 45, $\widehat{\rho }=-.15$, SE = .03, Z = −4.56, p < .0001, 95% CI = −.22 to −.09) and VoxBo (k = 4, $\widehat{\rho }=-.01$, SE = .19, Z = −.04, p = .97, 95% CI = −.38 to .37). Significant residual heterogeneity was observed in this analysis as well (Q_E[151] = 249.56, p < .0001).

2.6.6 |. Stereotactic normalization

Effect size outcomes were additionally moderated by stereotactic normalization choices (k = 154, Q_M[5] = 68.49, p < .0001) All post hoc analyses significantly differed (Q_M[2] ≥ 16.21, p ≤.0006) except the comparison between studies using MNI and those using Talairach templates (Q_M[2] = 3.73, p = .16). Studies that employed scanner-specific templates demonstrated the most negative outcomes (k = 10, $\widehat{\rho }=-.38$, SE = .08, Z = −4.57, p < .0001, 95% CI = −.55 to −.22), followed by studies using a group average (k = 12, $\widehat{\rho }=-.31$, SE = .08, Z = −4.03, p < .0001, 95% CI = −.47 to −.16), individual patient brains (k = 56, $\widehat{\rho }=-.16$, SE = .03, Z = −5.27, p < .0001, 95% CI = −.21 to −.10), MNI template (k = 72, $\widehat{\rho }=-.08$, SE = .04, Z = −1.93, p = .054, 95% CI = −.17 and .00), and Talairach template (k = 4, $\widehat{\rho }=-.01$, SE = .19, Z −.03, p = .98, 95% CI = −.37 to .36). Significant residual heterogeneity was also found (Q_E[149] = 231.07, p < .0001).

2.6.7 |. tfMRI asymmetry measure

tfMRI index also moderated effect size outcomes (k = 154, Q_M[2] = 49.33, p < .0001). Studies comparing ROI activation asymmetry yielded significantly larger outcomes (k = 12, $\widehat{\rho }=-.22$, SE = .08, Z = −2.57, p = .01, 95% CI = −.38 to −.05) than studies bearing an LI (k = 142, $\widehat{\rho }=-.16$, SE = .02, Z = −6.54, p < .0001, 95% CI = −.21 to −.11). However, significant residual heterogeneity resulted (Q_E[152] = 251.20, p < .0001).

2.7 |. Moderation analyses

2.7.1 |. Epilepsy focus

A linear mixed-effects model demonstrated the laterality of epilepsy focus significantly moderated effect size outcomes (Q_M[2] = 132.88, p < .0001). tfMRI studies of left TLE predicted postsurgical cognitive outcomes (k = 86, $\widehat{\rho }=-.27$, SE = .02, Z = −11.49, p < .0001, 95% CI = −.32 to −.23), whereas studies of right TLE did not (k = 68, $\widehat{\rho }=-.02$, SE = .03, Z = −.87, p = .38, 95% CI = −.08 to .03). In this analysis, residual heterogeneity remained (Q_E[152] = 198.56, p = .0067).

2.7.2 |. tfMRI domain

Investigation within studies of left TLE suggested effect size outcomes were further moderated by tfMRI domain (Q_M[2] = 74.34, p < .0001). Analysis showed language tfMRI studies were more robust in predicting postsurgical cognitive outcomes among these patients (k = 66, $\widehat{\rho }=-.27$, SE = .03, Z = −8.00, p < .0001, 95% CI = −.33 to −.20) as compared to memory tfMRI studies (k = 20, $\widehat{\rho }=-.27$, SE = .08, Z = −3.23, p = .0013, 95% CI = −.43 to −.11). This analysis also demonstrated significant residual heterogeneity (Q_E[84] = 198.56, p = .0067).

2.7.3 |. Cognitive domain

Effect size outcomes within language and memory tfMRI studies of left TLE were further moderated by postsurgical cognitive domain (Q_M[6] = 91.08, p < .0001).

Memory tfMRI studies of verbal memory outcomes (Figure 4A; k = 12, $\widehat{\rho }=-.37$, SE = .10, Z = −3.80, p = .0001, 95% CI = −.57 to −.18) yielded stronger prediction than studies relating language tfMRI with confrontation naming outcomes (Figure 3A; k = 35, $\widehat{\rho }=-.32$, SE = .05, Z = −6.41, p < .0001, 95% CI = −.41 to −.22), language tfMRI with verbal memory outcomes (Figure 3C; k = 21, $\widehat{\rho }=-.26$, SE = .05, Z = −5.64, p < .0001, 95% CI = −.35 to −.17), language tfMRI with verbal fluency outcomes (Figure 3B; k = 9, $\widehat{\rho }=-.19$, SE = .11, Z = −1.66, p = .097, 95% CI = −.41 to .03), memory tfMRI with nonverbal memory outcomes (Figure 4B; k = 8, $\widehat{\rho }=-.07$, SE = .14, Z = −.54, p = .59, 95% CI = −.35 to .20), and language tfMRI with nonverbal memory (Figure 3D; k = 1, $\widehat{\rho }=.21$, SE = .25, Z = .84, p = .40, 95% CI = −.28 to .69). No study included related memory tfMRI to confrontation naming or verbal fluency outcomes.

FIGURE 4.

Forest plot of associations of memory task-based functional magnetic resonance imaging with verbal memory (A) and nonverbal memory (B) outcomes in left temporal lobe epilepsy. CI, confidence interval

FIGURE 3.

Forest plot of associations of language task-based functional magnetic resonance imaging with confrontation naming (A), verbal fluency (B), verbal memory (C), and nonverbal memory (D) outcomes in left temporal lobe epilepsy. CI, confidence interval; NA, not applicable

All post hoc contrasts were significant (Q_M[2] ≥ 15.11, p ≤ .0003) except the comparisons between language tfMRI studies of nonverbal memory outcomes and language tfMRI studies of verbal fluency outcomes (Q_M[2] = 3.45, p = .53), memory tfMRI studies of nonverbal memory outcomes and language tfMRI studies of verbal fluency outcomes (Q_M[2] = 3.04, p = .53), and language tfMRI studies of nonverbal memory outcomes and memory tfMRI studies of nonverbal memory outcomes (Q_M[2] = .99, p = .61). Nonetheless, significant residual heterogeneity remained (Q_E[80] = 129.85, p = .0004).

2.7.4 |. Surgery type

Contrasting outcomes by surgery types within studies demonstrated significant moderation (Q_M[2] = 75.56, p < .0001). Studies of patients receiving left ATL were more robust (k = 66, $\widehat{\rho }=-.28$, SE = .03, Z = −8.12, p < .0001, 95% CI = −.35 to −.21) than studies of a mixed patient group (k = 20, $\widehat{\rho }=-.22$, SE = .07, Z = −3.10, p = .0019, 95% CI = −.36 to −.08). Significant residual heterogeneity remained in this analysis as well (Q_E[84] = 141.94, p < .0001).

2.8 |. Meta-regression analyses

2.8.1 |. Age

A linear mixed-effects analysis indicated no association between age and effect size outcomes within left TLE tfMRI studies (k = 61, Q_M[1] = .03, p = .87).

2.8.2 |. Sex

Sex, operationalized as the proportion in which the patient sample was male, was associated with more negative effect size outcomes in left TLE studies (Figure 5A; k = 84, Q_M[1] = 8.23, Z = −2.87, p = .004) and was shown to explain considerable variance in study outcomes (pseudo R² = 13.85%). Despite this association, significant heterogeneity remained (Q_E[82] = 127.39, p = .001).

FIGURE 5.

Fitted meta-regression scatterplot associating effect size outcomes in left temporal lobe epilepsy with birth sex (A), level of education (B), age at epilepsy onset (C), epilepsy duration (D), and normative language hemisphere dominance (E)

2.8.3 |. Education

Higher years of education was associated with more positive effect size outcomes in studies of left TLE (Figure 5B; k = 47, Q_M[1] = 27.46, Z = 5.24, p < .0001) and explained a large proportion of variance in study outcomes (pseudo R² = 100.00%). Interestingly, no significant heterogeneity remained (Q_E[45] = 43.81, p = .52).

2.8.4 |. Handedness

Handedness, defined as the proportion of the patient sample who were right-handed, did not explain significant variance in effect size outcome within studies of left TLE (k = 75, Q_M[1] = .53, p = .47).

2.8.5 |. Age at epilepsy onset

More advanced age at epilepsy onset among patients predicted more positive effect size outcomes within studies of left TLE (Figure 5C; k = 67, Q_M[1] = 12.75, Z = 3.57, p = .0004) and explained substantial variance in study outcomes (pseudo R² = 45.85%). Significant residual heterogeneity was present (Q_E[65] = 106.35, p = .0009).

2.8.6 |. Duration of epilepsy

Longer duration of epilepsy was associated with more negative effect size outcomes within studies of left TLE (Figure 5D; k = 64, Q_M[1] = 13.14, Z = −3.63, p = .0003) and explained substantial variance in study outcomes (pseudo R² = 55.23%). Significant residual heterogeneity was also demonstrated (Q_E[62] = 93.61, p = .0059).

2.8.7 |. Postsurgical seizure freedom

There was a trend-level relationship between the proportion of the sample who achieved seizure freedom after surgery and effect size outcomes within studies of left TLE (k = 47, Q_M[1] = 3.39, p = .07).

2.8.8 |. Presurgical intellectual functioning

Neither performance (nonverbal) intelligence quotient (IQ; k = 31, Q_M[1] = .21, p = .64) nor verbal IQ (k = 39, Q_M[1] = .02, p = .88) was related with effect size outcomes in left TLE studies. There was a trend-level relationship between general IQ and study outcomes; however, this association was not significant (k = 29, Q_M[1] = 2.76, p = .10).

2.8.9 |. Normative left hemisphere language dominance

More normative left hemisphere language dominance was associated with more negative effect size outcomes in studies of left TLE (Figure 5E; k = 55, Q_M[1] = 4.89, Z = −2.21, p = .027). The proportion of variance in effect size outcomes accounted for was substantial (pseudo R2 = 27.51%), although significant residual heterogeneity remained (Q_E[53] = 87.00, p = .002).

2.8.10 |. Postsurgical cognitive decline

Postsurgical cognitive decline among patients was not related to variances in effect size outcomes in studies of left TLE (k = 64, Q_M[1] = 1.83, p = .18).

2.9 |. Publication bias

Overall, publication bias analyses provided mixed evidence of bias. Visual inspection of a funnel plot showed right-sided asymmetry (Figure 2). Begg and Mazumdar rank correlation test was not significant (Τ = −.01, p = .90); however, Egger regression was significant (Z = 1.96, p = .049). Orwin fail-safe N suggested an additional 218 missing contradictory studies would be required to reduce the overall effect to a trivial result (i.e.,$\widehat{\rho }=-.05$). In light of mixed evidence of publication bias, Duval and Tweedie’s trim-and-fill approach was employed and imputed eight left-sided studies (SE = 4.24, p = .002), which significantly increased the magnitude of the overall effect (Figure S3;$\widehat{\rho }=-.18$, SE = .02, Z = −7.56, p < .0001, 95% CI = −.22 to −.13). Additional publication bias analyses were conducted within the literature on left TLE for the purposes of GRADE and are reported in Supplementary Material (p. 6–8; Table S3).

FIGURE 2.

Funnel plot

2.10 |. Certainty of evidence

Evidence tables are presented in Tables S2 and S3 for the reviewed literature in left TLE. The certainty of evidence for all bodies of evidence was rated as “low” or “very low” due to the observational nature of available studies, between-study heterogeneity, and publication bias. Notable strengths of the literature are that the level of bias and outcome measurement were appropriate to address the present review question.

3 |. DISCUSSION

The present findings further indicate tfMRI predicts postsurgical cognitive outcomes following temporal lobe epilepsy surgery; however, results were significant only for patients with left TLE. This differential result agrees with reports demonstrating patients with left TLE have more negative postsurgical cognitive outcomes,⁵¹ presumably because these patients receive intervention within the hemisphere that is likely dominant for language function.^2–6 As such, tfMRI may be more sensitive to postsurgical cognitive morbidity among this subgroup. Among patients with left TLE, language tfMRI was shown to be more broadly predictive of postsurgical cognitive outcomes than memory tfMRI, although memory tfMRI was more sensitive to verbal memory outcomes than language tfMRI. These effect estimates suggest memory tfMRI predicts approximately 12.5% of the variance in verbal memory outcome, whereas language tfMRI is associated with 9.6% and 6.5% of the variance in confrontation naming and verbal memory outcomes, respectively. Importantly, these effect sizes are within the small to moderate range according to effect size conventions,⁵² which suggest additional factors are involved. In general, these findings are concordant with the American Academy of Neurology practice guideline review that suggested the evidence base of memory tfMRI is more robust in the prediction of verbal memory outcomes as compared with the literature substantiating the forecasting of language outcomes based on language tfMRI.² Interestingly, these results also agree with a recent meta-analysis that demonstrated memory IAP has reduced sensitivity but greater specificity with respect to memory outcomes relative to language IAP.⁷ Overall, the present meta-analysis further substantiates the utility of tfMRI in epilepsy surgery and provides the first effect size estimates associating tfMRI lateralization with postsurgical cognitive outcomes among patients with left TLE. These estimates can be used to quantitate the probability of poor postsurgical cognitive outcomes following epilepsy surgery based on tfMRI in conjunction with other diagnostic modalities.

Moderation analyses demonstrated greater sensitivity to less favorable outcomes among male patients, patients with lower educational attainment, patients of a younger age at epilepsy onset, patients with a longer disease course, and lastly, patients with normative left hemisphere language dominance. The latter finding is concordant with the literature showing that postsurgical morbidity is more likely in the setting of intervention in the dominant hemisphere.^2,7,53,54 These results also show tfMRI is sensitive to the potential neurodegenerative effects of intractable seizures in patients with longer histories of epilepsy.⁵⁵ However, the relationship between tfMRI–outcome associations and age at epilepsy onset may be in the direction opposite that expected, as prior evidence shows epilepsy onset at an earlier age is associated with fewer declines following surgery.^53–56 The finding with respect to educational attainment may be related to reduced cognitive reserve among patients with lower education.⁵⁷ Why tfMRI is more sensitive to worse outcomes in male patients is unclear; this may be due to sampling biases in the literature. Finally, outcome prediction was also more robust for studies of solely ATL as compared to studies involving more heterogeneous patient groups, which is a finding that likely relates to differences in efficacy and postsurgical morbidity between distinct TLE surgical interventions.^58,59 However, the larger pool of data available for ATL as opposed to other procedures may account for this.² The breadth of moderators noted among these analyses underscores that the prediction of postsurgical cognitive outcomes in left TLE using tfMRI is multifactorial in nature.

Apparent among these findings was that methodological and reporting practices influenced results as well. The significant residual heterogeneity noted among the moderation analyses suggests that between-study variance is considerable, and variables of interest do not fully account for study outcomes. The results may have been affected by between-study heterogeneity. This heterogeneity suggests a need for methodological harmonization among tfMRI studies predicting postsurgical cognitive outcomes. The sensitivity analyses showed effect sizes varied as a function of follow-up interval, reported effect size, scanner magnetic field strength, tfMRI analysis software, stereotactic normalization scheme, and tfMRI asymmetry measure. Specifically, the sensitivity analyses demonstrated study outcomes are affected by study differences beyond those normally expected (e.g., sampling or measurement error or disease-relevant characteristics). Because tfMRI has become common in epilepsy surgery, the impetus for methodological harmonization herein also extends to clinical practice.

3.1 |. Limitations and future directions

Our review should be considered in the context of its limitations. First, our literature search may have been impacted by our secondary citation review of included articles. This may have impacted the systematic nature of our literature search strategy. Second, although no formal eligibility criteria concerned age of patients, our review included only studies of adults. Future studies should consider pediatric samples. Third, the estimate of language tfMRI studies predicting nonverbal memory outcomes in left TLE was limited by a low number of studies. Future research should examine whether language tfMRI may be useful to forecast nonverbal outcomes, especially as normative language function may eventuate in lesser postsurgical morbidity among nonverbal abilities and vice versa. Fourth, psychometric differences between cognitive tests may have influenced results due to their variable sensitivities to specific cognitive functions. In service of methodological and clinical harmonization, future studies should examine whether differences between cognitive tests influence prediction of postsurgical cognitive outcomes. Fifth, many of our planned analyses investigating the contribution of certain moderators to study outcomes were limited by insufficient reporting within studies. Transparent reporting of study characteristics is crucial to explaining heterogeneity within a meta-analysis, particularly when significant heterogeneity is observed, such as in the present setting. Sixth, despite attempts to contact select authors, several studies had to be excluded due to issues regarding their methodology, study reporting, or both. This further underscores a need for methodological standardization and rigorous study reporting. Finally, as our findings show the multifactorial nature of tfMRI for the prediction of postsurgical cognitive outcome, multivariate meta-analysis may advance this literature, given its potential to examine multiple predictors simultaneously.⁶⁰ However, sufficient studies among a rather homogeneous literature would be needed to permit such an undertaking.

4 |. CONCLUSIONS

Our findings indicate tfMRI is a modest predictor of postsurgical cognitive outcomes in left TLE. Importantly, language tfMRI was more broadly predictive of outcomes than memory tfMRI, whereas memory tfMRI was the strongest predictor of verbal memory outcomes. Our meta-analysis provides the first effect size estimates summarizing the predictive validity of tfMRI in the setting of epilepsy surgery. We have also investigated the demographic, clinical, methodological, and reporting practices that may influence the sensitivity of tfMRI to postsurgical cognitive outcomes. Overall, these results demonstrate that tfMRI should be considered in the context of a larger surgical workup and underscore the added value of tfMRI. Our findings may also serve as an impetus for study harmonization.

Key Points.

• Language tfMRI more broadly predicts postsurgical cognitive outcomes in left-sided temporal lobe epilepsy than memory tfMRI.

• Memory tfMRI is a more robust predictor of verbal memory outcomes than language tfMRI, particularly in left-sided temporal lobe epilepsy.

• tfMRI did not consistently predict postsurgical cognitive outcomes in right-sided temporal lobe epilepsy.

• Several factors moderate the sensitivity of tfMRI to postsurgical cognitive outcomes in left-sided temporal lobe epilepsy.

• Methodological and reporting practices influenced study outcomes, suggesting a need for study harmonization.