MTL seeds
To ensure unbiased ROI selection for group comparisons, seed selection was based on AAL-based anatomical masks (WFU-PickAtlas toolbox v3.0). Healthy controls had the left and right hippocampi as MTL seeds, based on their role in successful memory.3,17 For patients, one MTL seed included the non-resected contralesional hippocampus. Ipsilesionally, the remnant MTL seed entailed remnant hippocampus and parahippocampal gyrus, based on left and right ATLR group resection masks39,48. All participants underwent standard ATLR with little within-group difference in the size of resection cavity.
Generalized PPI analysis
A generalized form of PPI (gPPI) spanned the entire experimental space,35 modelling beta-estimates of all subsequent memory conditions (further details in supplementary material).
At the participant-level, seed-to-whole-brain connectivity analysis was conducted on MATLAB.35 For verbal and visual memory separately, the subject-level gPPI model included three regressors: time-course of each event-related task-condition, timeseries of one MTL seed, and of the PPI term (task*seed interaction). All six task-conditions were modelled to better probe successful verbal and visual memory effect.35 The average seed time-course was extracted within each anatomical MTL mask/seed. Physiological and psychological variables were treated as nuisance regressors. T-contrasts were generated for each MTL seed and words/faces subsequently remembered, revealing whole-brain cortical areas significantly more correlated with the seed during successful memory encoding than during uncertain/failed conditions, based on the PPI term prediction35.
For each participant, separate GLMs were performed for each MTL seed. The same task and seed regressors were used for all participants and at 3–12-month and 10-year follow-ups. Resulting single-level gPPI t-contrasts of successful subsequent memory were used for group-level random-effects analyses (see below section).
Statistical analyses
Clinical and neuropsychological data
Data was analysed using R 4.0.5. Demographics were evaluated using Fisher’s exact test for sex proportion, one-way ANOVA for parametric z-scores (memory and IQ), and Kruskal–Wallis tests for nonparametric continuous variables (age, ILAE outcomes, ASM intake and change).42 For multiple comparisons, post-hoc tests were corrected using Tukey's HSD adjustment.
Longitudinal assessment of the functional memory network
All data was analysed with SPM12.
Long-term changes in functional connectivity from 3–12-month to 10-year follow-ups
Mixed ANOVAs, using flexible factorial design with IQ as confound regressor, were performed to investigate changes in MTL-seeded memory connectivity between short-term and long-term follow-ups in the individual patient groups compared to changes in test-retest over that timeline in healthy subjects. A distinct analysis was conducted for faces and words successfully remembered, seeding from each MTL separately.
For each subject, the relevant first-level gPPI t-contrasts for each of the scanning timepoint (3–12 months and 10 years) were entered. Each flexible factorial design involved a random subject factor, a three-levels group factor (controls, left and right ATLR), and a two-levels condition factor (short-term and long-term postoperative scans), allowing the investigation of a Group x Condition interaction for the successful memory contrasts. Differences in activations across scanning sessions were compared between ATLR groups and controls in t-contrasts: 10-year connectivity > 3–12-month connectivity in left or right ATLR versus controls, and 10-year connectivity < 3–12-month connectivity in left or right ATLR versus controls.
In summary, flexible factorial t-contrasts modelled in each ATLR group, the within-subject differences in MTL-seeded, whole-brain functional connectivity across postoperative follow-ups, beyond the connectivity changes seen in controls.
Correlation between functional reorganization in TLE and memory recovery
Three-way ANCOVAs were conducted for each MTL seed and each successful memory contrast (words/faces), to investigate which differences in functional connectivity from 3–12-month and 10-year follow-ups were related with improvement in memory functions over this timeline. Positive correlations were examined using BMIPB (I or II) verbal and visual learning scores converted into age-normalized z-scores. Difference in z-scores between 3–12-month and 10-year follow-ups was used as continuous variables in three-ways ANCOVAs.
Statistical thresholds
Given our a priori hypothesis of increased local MTL connectivity (including the fusiform gyrus),5,9,17 MTL connectivity was corrected for multiple comparisons at P < 0.05 voxel-wise, controlling for family-wise error rate via small volume correction.5,32,49 This included a 6 mm-radius sphere in contralesional MTL regions,3,12 and 3 mm-radius sphere in remnant hippocampus and parahippocampal gyrus to avoid resection cavity-related activation. All reported seed-to-remnant hippocampus/parahippocampus connectivity was validated against artifacts using exclusive MTL group-resection masks.
Group comparison and correlation analyses generated highly specific MTL-to-whole-brain, longitudinal, event-related t-contrasts.50 Thus, at the extra-MTL level, functional connectivity is reported at an exploratory P < 0.001 threshold (uncorrected), alike previous longitudinal, event-related and network fMRI studies.5,12,49