Patient Selection and Data Acquisition

Patients were selected from the Epileptology database of Timone Hospital, Marseille, if they had: video-SEEG-recorded seizures with testing-proven postictal aphasia, pre/post-implantation cerebral computed tomography-magnetic resonance imaging (CT-MRI) data and SEEG confirmation of hemispheric language dominance.

Presurgical evaluations included detailed medical history, neurological examination, neurocognitive testing, cerebral fluorodeoxyglucose-positron emission tomography (FDG-PET) and MRI scans and long-term scalp-video-EEG and SEEG monitoring. Additional assessments (eg, language fMRI, Wada testing, cerebral 7T MRI, and magnetoencephalography [MEG]) were conducted on a case-by-case basis. Demographic, clinical, and follow-up data were retrieved from medical records. All patients provided informed written consent and the study was approved by the ethics committee of Assistance Publique – Hôpitaux de Marseille (PADS VNMG2V2024).

SEEG Procedure

SEEG explorations were conducted in accordance with the French national guidelines,¹² using 10 to 18-contact intracerebral electrodes (Dixi or Alcis; 2 mm contact length, 1.5 mm spacing, and 0.8 mm diameter) implanted based on the noninvasive anatomic-electro-clinical EZN hypothesis. Post-implantation cerebral CT and/or MRI scans were conducted to verify electrode placement accuracy and identify any periprocedural complications. Signal recording was performed on a 128 or 256-channel Natus system, with a sampling rate of 512 or 1,024 hertz (Hz). Hardware filtering included a high-pass filter (1 Hz cutoff, −3 decibels [dB]) and an anti-aliasing low-pass filter (170 Hz or 340 Hz cutoff ).

SEEG Signal Analysis

All signal analyses were performed in AnyWave software¹³ in bipolar montage. Any channels showing artifacts were excluded based on visual inspection. The EZN, the propagation-zone network (PZN) and the non-involved areas (NIZ) were defined by 3 expert clinicians (authors I.F.B., F.B., and A.T.), based on visual and quantitative SEEG-signal analysis using the Epileptogenicity Index (EI) and the Connectivity Epileptogenicity Index (cEI).^{14, 15} The EI measures epileptogenicity based on the energy ratio between high-frequency and low-frequency bands and the delay of the abrupt frequency transition in a specific brain structure relative to the first structure involved by the fast discharge. The cEI integrates the EI with a measure of directed functional connectivity (“out-degree”), allowing the evaluation of the leading nodes in the network.

Permutation Entropy Method

PE quantifies the probability distribution of ordinal patterns, which are determined by the rank of values sampled at equidistant points in a time series.¹⁶ Starting with a one-dimensional time series, vectors are created based on the embedding dimension D, which determines the number of samples per vector and a fixed time delay τ, for sample selection. Each vector is assigned an ordinal pattern π, corresponding to the rank of sample amplitudes. The probability p(π) of each ordinal pattern occurring within a given time window is then computed. The entropy within the window is defined as $$$ H(D)=-\frac{1}{{\mathit{\log}}_2D!}\sum p\left(\pi \right){\mathit{\log}}_2\left(p\left(\pi \right)\right) $$$. The values of H(D) range from 0 to 1, where smaller values indicate a more regular and deterministic time series and values closer to 1 indicate a noisier, more random time series. SEEG PE analysis was performed using an in-house Matlab (Mathworks, Natick, MA) plug-in (details in El Youssef et al).¹⁷ Based on exploratory testing, review of the literature, and our previous work, the PE analysis parameters used were D = 3 samples and τ = 1 sample (consecutive observations), using a 5-second sliding time window with a 2.5-second overlap.^{9, 10, 18}

Definition of Regions of Interest

To accurately map SEEG contacts to brain regions, GARDEL software was used.¹⁹ It co-registers the pre-implantation cerebral T1-MRI with the post-implantation cerebral CT images, followed by semi-automatic recognition and anatomic localization of each electrode contact, based on the Virtual Epileptic Patient (VEP) atlas²⁰ parcellation of the patient's cerebral MRI scan.

Network and State Assessment

A thorough visual analysis of (peri)ictal SEEG recordings was performed. Seizure onset was defined as the first SEEG change within a sustained rhythmic discharge and seizure offset as the cessation of ictal activity in the last involved bipolar derivation. Contacts were classified as EZN (EI ≥ 0.4 and/or cEI ≥ 0.65), PZN (0.1 < EI < 0.4 and/or 0.3 < cEI < 0.65) or NIZ (all other contacts).²¹ False cEI positives values, such as those arising from physiological fast activity (eg, motor cortex sampling or during cortical arousal), were identified and excluded from cEI computation following expert visual inspection. Comprehensive clinical assessment and SEEG trace inspection were critical for final classification, beyond quantified biomarkers. Brain regions were classified using the maximal cEI value when sampled by multiple contacts.

PE values were computed for the (peri)ictal periods of spontaneous and stimulation-induced seizures. As per our previous study,¹⁰ we considered that the SEEG signal returned to its preictal complexity (“come-back point”) when for 25 consecutive seconds onward (10 × 2.5-second overlap) the mean PE value returned within the preictal 2 SD baseline interval. The period between seizure offset and the come-back point was named Postictal Alteration Time (PAT). The PAT was quantified globally (mean across all channels; Fig 1A), hemispherically (mean per hemisphere; Fig 1B) and regionally (maximum PAT value from all channels sampling a specific cerebral area).

Language Organization Exploration

Hemispheric language dominance was determined on a case-by-case basis using both noninvasive and invasive assessments.⁶ The noninvasive explorations included: (1) video-EEG electro-clinical ictal and postictal correlations, (2) clinical data (eg, handedness), (3) age-appropriate neuropsychological testing (eg, the Wechsler Adult Intelligence Scale and the Wechsler Intelligence Scale for Children), and (4) language fMRI. Invasive assessments involved: (1) video-SEEG electro-clinical correlations during ictal and postictal periods, (2) SEEG recording of auditory cortex evoked potentials in response to French-language voiced (/ba/) and voiceless (/pa/) stop consonants, (3) functional mapping of the language eloquent cortex using direct electrical cortical stimulations and identification of regions producing gamma activity during a picture naming task, and (4) WADA testing.

Postictal Aphasia Bedside Evaluation

Postictal language assessment was performed at the time of the video-SEEG exploration using an in-house test battery (Supplementary Fig S1). It addressed the following tasks: naming, automatic speech (eg, reciting days/months/seasons or counting), repetition (eg, 2 words of varying complexity and a complete phrase), verbal and written comprehension (eg, simple and complex commands), reading aloud (eg, simple words, phrases and logatomes), phonemic and semantic verbal fluency (eg, words starting with “L” and naming fruits), and working memory (eg, 3-word recall and simple mental arithmetic).

Electro-Anatomic-Clinical Postictal Aphasia Analysis

Video-SEEG clinical and electrical data were first analyzed separately and independently. Postictal clinical language evaluations were retrospectively carried out by author Ionuț-Flavius Bratu using ictal-postictal video-SEEG recordings. As per standard practice in our department, patients underwent postictal language testing within 3 defined time intervals: 0 to 5, 5 to 10, and 10 to 15 minutes after seizure termination. For finer electro-anatomic-clinical correlations, each tested language domain was assigned to a 2.5-minute subinterval within the 0 to 15 minutes postictal window (0–2.5, 2.5–5, 5–7.5, 7.5–10, 10–12.5, and 12.5–15 minutes; Fig 2). Each language domain was scored as 0 (no recovery), 1 (partial recovery), or 2 (full recovery). Postictal brain region complexity recovery (PAT per region) was quantified according to the previously described considerations and classified according to the 6 postictal time intervals.

To assess the relationship between postictal language domain recovery and entropy recuperation in distinct left-hemispheric regions of SEEG-confirmed left-hemisphere language-dominant patients (see Fig 2), we used a comprehensive approach that included region selection based on frequency of occurrence across seizures (eg, due to SEEG-sampling), statistical resampling and permutation testing. Brain regions sampled in at least 12 seizures were retained, even if all 12 seizures came from a single patient, to balance statistical power with adequate representation of less frequently sampled areas.

Co-occurrences of the entropy and language domain recovery for each brain region were computed across the six 2.5-minute intervals. We applied a permutation-based resampling method to assess the significance of observed co-occurrences. In each of 1,000 permutations, shuffling was performed at multiple levels: first, the language recovery intervals were randomly shuffled within each patient, then within each seizure, and finally the patients were shuffled as an “en bloc” group (stratified by the number of seizures). This generated a null distribution of match probabilities between shuffled language recovery intervals and the original entropy recovery intervals. From this null distribution, a 95th percentile threshold was derived as the cutoff for statistical significance. The observed proportion of co-occurrences, defined as the fraction of matching intervals between language recovery and entropy recovery, was evaluated against this threshold. Regions where the observed proportion exceeded the threshold were considered to show statistically significant electro-clinical correlations, entropy-language domain recovery.

Ventral-Dorsal Stream Electro-Clinical Framework

Stream dichotomization was assessed from an electrical-clinical point of view: stream PAT recovery and language recuperation. Using the VEP atlas parcellation, we classified regions in ventral and/or dorsal language stream areas based on a literature review^{3, 22} (Supplementary Table S1). For each stream, PAT was computed from the highest values of the individual brain regions. We further analyzed a subgroup of seizures in which 3 specific language abilities were tested in the same 2.5-minute interval at the beginning of the assessment and whose recovery could be traced reliably across the intervals: (1) reading aloud (reflecting both streams, dorsal more than ventral), (2) written comprehension (for the ventral stream), and (3) repetition (for the dorsal stream). This analysis was carried out in left hemisphere regions in SEEG-confirmed left-hemisphere language-dominant patients.

Postictal Aphasia Scale

To assess inter-rater agreement, postictal language evaluations were performed on a subset of seizures by 2 assessors (author I.F.B. and the speech therapy team: authors L.S. and F.M.), using the same scoring criteria as previously described. Language performance across relevant domains was evaluated at 2.5-minute intervals during the postictal period, based solely on bedside video recordings. Although author Ionuț-Flavius Bratu was aware of the inclusion criteria for patient selection, all video analyses were performed blinded to other clinical and electrophysiological information. The second rater, independently re-scored the same seizures, fully blinded to electro-clinical data. Inter-rater agreement was evaluated using the quadratic weighted kappa test, which assigns higher penalties for larger rating discrepancies. Language functions that were frequently tested and showed high bedside feasibility and agreement were selected to create a simplified postictal aphasia bedside scale. This scale was applied to re-analyze data at 5-minute intervals and its scores were assessed in relationship to left-hemisphere regional PAT in SEEG-confirmed left-hemisphere language-dominant patients.

Statistical Analysis

Statistical analyses were conducted using GraphPad Prism (version 10.3.1), MedCalc Software (version 22.026), Matlab (version 2023b), and RStudio (version 4.3.2) for Windows. The Shapiro–Wilk normality test guided the choice of parametric versus nonparametric tests (threshold p < 0.05). To address multiple comparisons across several brain regions, the false discovery rate (FDR) correction was applied (Benjamini–Hochberg Method), Matlab (version 2023b) and RStudio (version 4.3.2) for Windows.

For patients who experienced both seizure types (with and without postictal aphasia), PAT values were averaged per seizure type and compared using the Wilcoxon signed-rank test. Furthermore, linear mixed-effects models were fitted using the lmerTest package in R software to account for repeated measures in patients with multiple seizures and relevant demographic and clinical covariates. The outcome variable was PAT (global or at dominant hemisphere level). Postictal aphasia (present/absent) was included as the main fixed effect of interest, with seizure type (spontaneous/stimulation-induced), age group (pediatric = < 18 years old and adults = ≥ 18 years old) and EZN-language overlap (complete, partial, or none) as additional covariates. A random intercept accounted for repeated seizures per patient. Similar models were applied to study the Postictal Aphasia Score (PAS) score (at different postictal time points) as the outcome variable and regional PAT as the main fixed effect, controlling for the same covariates. For network-level PAT analysis, the maximum PAT per region was first identified, then averaged across the regions within each network (EZN, PZN, or NIZ) and across seizures for each patient. For comparisons between PAT values at the network level, we used Kruskal-Wallis test with post hoc correction (Dunn's test). To account for repeated measures (potentially several seizures per patients), a linear mixed-effects model was also fitted with the network as a fixed effect and the patient as random effect.

Cohort Characteristics

Patients’ characteristics are summarized in Table 1 and individual patient-specific information are provided in Supplementary Table S2. We analyzed 322 focal seizures (without bilateral to tonic–clonic evolution) from 98 patients. Postictal aphasia occurred in 237 seizures (31 stimulation-induced and 206 spontaneous). Eighty-five seizures were not followed by aphasia (25 stimulation-induced and 60 spontaneous). The mean age at SEEG was 32.21 years (range = 6–61 years). Seventy percent of the cases had temporal EZN, 18% were temporal plus, and 12% were extra-temporal. EZN was left-sided in 60% of patients, right-sided in 16%, and bilateral in 24%. Post-SEEG curative-aim epilepsy surgery was performed in 52% of patients, with an Engel I outcome at last follow-up achieved in 63%. The most common histopathological finding was focal cortical dysplasia (25%), whereas 29% of the resected tissue samples showed no specific pathology. SEEG language assessment revealed typical left hemisphere language organization in 84 patients (280 seizures), atypical bilateral organization in 12 cases, and atypical right hemisphere organization in 2 cases.

Postictal Electro-Clinical Correlations

Global PAT was statistically significantly higher in seizures followed by postictal deficit compared with those without postictal aphasia (Wilcoxon signed-rank test, p < 0.0001), both in the entire seizure cohort (spontaneous and stimulation-induced, effect size −0.719) and for spontaneous seizures alone (effect size = −0.716; Fig 3A). This was also true for left hemisphere PAT in patients with SEEG-confirmed left-hemispheric language dominance (Wilcoxon signed-rank test, p < 0.001; spontaneous and stimulation-induced seizures, effect size = −0.713 and spontaneous seizures alone, effect size = −0.688; Fig 3B). Furthermore, PAT revealed a hierarchical distinction among networks: EZN > PZN > NIZ (Fig 3C), with a longer PAT in more epileptogenic regions (Kruskal-Wallis analysis with post hoc Dunn's test; EZN vs. NIZ, p < 0.0001, effect size = 0.451; EZN vs. PZN, p = 0.0165, effect size = 0.198 and PZN vs. NIZ, p = 0.0012, effect size = 0.253). These results were confirmed (p < 0.05) by mixed-effects models accounting for both repeated measurements for patients with multiple seizures and relevant patient demographic and clinical features.

Co-occurrences between SEEG signal complexity recovery and language-domain recuperation are summarized for the 10 highest match proportions in Supplementary Table S3, with significant co-occurrences identified through permutation-based resampling illustrated in Figure 4. For naming, significant co-occurrences were observed with entropy recovery in pars opercularis, rostral superior frontal sulcus, and anterior middle temporal gyrus. Automatic speech recovery was significantly associated with complexity recovery in the anterior middle temporal gyrus, precuneus, and putamen. Repetition recovery showed significant co-occurrences with regional complexity recovery in the posterior inferior temporal and lingual sulci, posterior cingulate cortex, precuneus, and putamen. Significant verbal comprehension-entropy recuperation co-occurrences were found for the caudate nucleus, insular gyri brevi, posterior inferior and anterior superior temporal sulci, putamen, planum polare, posterior middle temporal, and anterior inferior temporal gyri. For written comprehension, significant co-occurrences were shown with complexity recovery in the amygdala, angular gyrus, pars opercularis and triangularis, inferior frontal sulcus, insular gyri brevi, and superior parietal lobule. Reading aloud-entropy recovery co-occurred significantly in the angular, superior frontal and caudal middle frontal gyri, pars opercularis, insular gyri brevi, parietal operculum, rostral superior frontal sulcus, and superior parietal lobule. Phonemic verbal fluency recovery was linked to complexity recovery in the insular gyri longi, whereas semantic verbal fluency recovery highlighted significant co-occurrences with entropy return to baseline in the pars opercularis. Encoding and recall recovery significantly co-occurred with complexity return in the posterior supramarginal gyrus, gyrus of Heschl, parahippocampal cortex, and collateral sulcus. For mental arithmetic, significant associations were identified with complexity recovery in the occipito-temporal and posterior superior temporal sulci.

We identified 9 seizures with postictal aphasia from 8 SEEG-confirmed left-hemisphere language-dominant patients where reading aloud, written comprehension, and repetition were tested, as outlined in the Materials and Methods. In all cases, ventral stream PAT was longer than dorsal stream PAT. Clinically, patients regained the ability to repeat and read aloud before written comprehension (Table 2).

Postictal Aphasia Scale

Seventy-five seizures were used to evaluate the inter-rater agreement of postictal clinical language evaluations. Substantial inter-rater agreement (Cohen κ = 0.61–0.8) was found for working memory, phonemic verbal fluency, verbal comprehension, and reading aloud tasks. For all the other language functions, inter-rater agreement was very high (Cohen κ > 0.81), the highest being for automatic speech (Cohen κ = 0.96). Furthermore, we developed the PAS, assessing verbal and written comprehension, reading aloud, naming, automatic speech, and repetition, details of which can be found in Table 3. It is based on our in-house postictal beside testing battery (see Supplementary Fig S1), partially published in its ictal form in Trebuchon et al.²³ Each item is scored according to the patient's ability to perform a task: 0 (not able), 1 (partially able), and 2 (fully able). PAS scores range from 0 (no recovery, complete aphasia) to 12 (full recovery, no aphasia) and revealed Cohen's κ values of 0.97, 0.97, and 0.91 at 5, 10, and 15 minutes, respectively. A significant relationship (p < 0.05 corrected) was found between PAT of certain left-hemispheric brain regions and PAS. At 5 minutes, significant regions were widely distributed across the lateral frontal, insular, and mesial parietal cortex and the temporal lobe. The highest model estimated coefficients were observed for the anterior inferior temporal gyrus, anterior superior temporal sulcus, rhinal cortex, and inferior frontal sulcus. At 10 minutes, significant associations involved the gyrus of Heschl, parietal operculum, and the lateral-posterior part of the superior temporal gyrus. This latter region remained significant at 15 minutes, along with the orbitofrontal cortex (Fig 5A). At 5 and 10 minutes, PAS language domains indicated a global aphasia pattern, characterized by production deficits and a gradual resolution of comprehension deficits. By 15 minutes, the aphasia pattern shifted to subtle production-related impairments, notably in naming (Fig 5B).