Artificial intelligence in epilepsy - applications and pathways to the clinic.

Artificial intelligence (AI) is rapidly transforming health care, and its applications in epilepsy have increased exponentially over the past decade. Integration of AI into epilepsy management promises to revolutionize the diagnosis and treatment of this complex disorder. However, translation of AI into neurology clinical practice has not yet been successful, emphasizing the need to consider progress to date and assess challenges and limitations of AI. In this Review, we provide an overview of AI applications that have been developed in epilepsy using a variety of data modalities: neuroimaging, electroencephalography, electronic health records, medical devices and multimodal data integration. For each, we consider potential applications, including seizure detection and prediction, seizure lateralization, localization of the seizure-onset zone and assessment for surgical or neurostimulation interventions, and review the performance of AI tools developed to date. We also discuss methodological considerations and challenges that must be addressed to successfully integrate AI into clinical practice. Our goal is to provide an overview of the current state of the field and provide guidance for leveraging AI in future to improve management of epilepsy.

Utility of intracranial EEG networks depends on re-referencing and connectivity choice.

Studies of intracranial EEG networks have been used to reveal seizure generators in patients with drug-resistant epilepsy. Intracranial EEG is implanted to capture the epileptic network, the collection of brain tissue that forms a substrate for seizures to start and spread. Interictal intracranial EEG measures brain activity at baseline, and networks computed during this state can reveal aberrant brain tissue without requiring seizure recordings. Intracranial EEG network analyses require choosing a reference and applying statistical measures of functional connectivity. Approaches to these technical choices vary widely across studies, and the impact of these technical choices on downstream analyses is poorly understood. Our objective was to examine the effects of different re-referencing and connectivity approaches on connectivity results and on the ability to lateralize the seizure onset zone in patients with drug-resistant epilepsy. We applied 48 pre-processing pipelines to a cohort of 125 patients with drug-resistant epilepsy recorded with interictal intracranial EEG across two epilepsy centres to generate intracranial EEG functional connectivity networks. Twenty-four functional connectivity measures across time and frequency domains were applied in combination with common average re-referencing or bipolar re-referencing. We applied an unsupervised clustering algorithm to identify groups of pre-processing pipelines. We subjected each pre-processing approach to three quality tests: (i) the introduction of spurious correlations; (ii) robustness to incomplete spatial sampling; and (iii) the ability to lateralize the clinician-defined seizure onset zone. Three groups of similar pre-processing pipelines emerged: common average re-referencing pipelines, bipolar re-referencing pipelines and relative entropy-based connectivity pipelines. Relative entropy and common average re-referencing networks were more robust to incomplete electrode sampling than bipolar re-referencing and other connectivity methods (Friedman test, Dunn-Sidák test P < 0.0001). Bipolar re-referencing reduced spurious correlations at non-adjacent channels better than common average re-referencing (Δ mean from machine ref = -0.36 versus -0.22) and worse in adjacent channels (Δ mean from machine ref = -0.14 versus -0.40). Relative entropy-based network measures lateralized the seizure onset hemisphere better than other measures in patients with temporal lobe epilepsy (Benjamini-Hochberg-corrected P < 0.05, Cohen's d: 0.60-0.76). Finally, we present an interface where users can rapidly evaluate intracranial EEG pre-processing choices to select the optimal pre-processing methods tailored to specific research questions. The choice of pre-processing methods affects downstream network analyses. Choosing a single method among highly correlated approaches can reduce redundancy in processing. Relative entropy outperforms other connectivity methods in multiple quality tests. We present a method and interface for researchers to optimize their pre-processing methods for deriving intracranial EEG brain networks.

Disparities in seizure outcomes revealed by large language models.

OBJECTIVE: Large-language models (LLMs) can potentially revolutionize health care delivery and research, but risk propagating existing biases or introducing new ones. In epilepsy, social determinants of health are associated with disparities in care access, but their impact on seizure outcomes among those with access remains unclear. Here we (1) evaluated our validated, epilepsy-specific LLM for intrinsic bias, and (2) used LLM-extracted seizure outcomes to determine if different demographic groups have different seizure outcomes. MATERIALS AND METHODS: We tested our LLM for differences and equivalences in prediction accuracy and confidence across demographic groups defined by race, ethnicity, sex, income, and health insurance, using manually annotated notes. Next, we used LLM-classified seizure freedom at each office visit to test for demographic outcome disparities, using univariable and multivariable analyses. RESULTS: We analyzed 84 675 clinic visits from 25 612 unique patients seen at our epilepsy center. We found little evidence of bias in the prediction accuracy or confidence of outcome classifications across demographic groups. Multivariable analysis indicated worse seizure outcomes for female patients (OR 1.33, P ≤ .001), those with public insurance (OR 1.53, P ≤ .001), and those from lower-income zip codes (OR ≥1.22, P ≤ .007). Black patients had worse outcomes than White patients in univariable but not multivariable analysis (OR 1.03, P = .66). CONCLUSION: We found little evidence that our LLM was intrinsically biased against any demographic group. Seizure freedom extracted by LLM revealed disparities in seizure outcomes across several demographic groups. These findings quantify the critical need to reduce disparities in the care of people with epilepsy.

Mapping hippocampal and thalamic atrophy in epilepsy: A 7-T magnetic resonance imaging study.

OBJECTIVE: Epilepsy patients are often grouped together by clinical variables. Quantitative neuroimaging metrics can provide a data-driven alternative for grouping of patients. In this work, we leverage ultra-high-field 7-T structural magnetic resonance imaging (MRI) to characterize volumetric atrophy patterns across hippocampal subfields and thalamic nuclei in drug-resistant focal epilepsy. METHODS: Forty-two drug-resistant epilepsy patients and 13 controls with 7-T structural neuroimaging were included in this study. We measured hippocampal subfield and thalamic nuclei volumetry, and applied an unsupervised machine learning algorithm, Latent Dirichlet Allocation (LDA), to estimate atrophy patterns across the hippocampal subfields and thalamic nuclei of patients. We studied the association between predefined clinical groups and the estimated atrophy patterns. Additionally, we used hierarchical clustering on the LDA factors to group patients in a data-driven approach. RESULTS: In patients with mesial temporal sclerosis (MTS), we found a significant decrease in volume across all ipsilateral hippocampal subfields (false discovery rate-corrected p [pFDR] < .01) as well as in some ipsilateral (pFDR < .05) and contralateral (pFDR < .01) thalamic nuclei. In left temporal lobe epilepsy (L-TLE) we saw ipsilateral hippocampal and some bilateral thalamic atrophy (pFDR < .05), whereas in right temporal lobe epilepsy (R-TLE) extensive bilateral hippocampal and thalamic atrophy was observed (pFDR < .05). Atrophy factors demonstrated that our MTS cohort had two atrophy phenotypes: one that affected the ipsilateral hippocampus and one that affected the ipsilateral hippocampus and bilateral anterior thalamus. Atrophy factors demonstrated posterior thalamic atrophy in R-TLE, whereas an anterior thalamic atrophy pattern was more common in L-TLE. Finally, hierarchical clustering of atrophy patterns recapitulated clusters with homogeneous clinical properties. SIGNIFICANCE: Leveraging 7-T MRI, we demonstrate widespread hippocampal and thalamic atrophy in epilepsy. Through unsupervised machine learning, we demonstrate patterns of volumetric atrophy that vary depending on disease subtype. Incorporating these atrophy patterns into clinical practice could help better stratify patients to surgical treatments and specific device implantation strategies.

iEEG-recon: A fast and scalable pipeline for accurate reconstruction of intracranial electrodes and implantable devices.

OBJECTIVE: Clinicians use intracranial electroencephalography (iEEG) in conjunction with noninvasive brain imaging to identify epileptic networks and target therapy for drug-resistant epilepsy cases. Our goal was to promote ongoing and future collaboration by automating the process of "electrode reconstruction," which involves the labeling, registration, and assignment of iEEG electrode coordinates on neuroimaging. We developed a standalone, modular pipeline that performs electrode reconstruction. We demonstrate our tool's compatibility with clinical and research workflows and its scalability on cloud platforms. METHODS: We created iEEG-recon, a scalable electrode reconstruction pipeline for semiautomatic iEEG annotation, rapid image registration, and electrode assignment on brain magnetic resonance imaging (MRI). Its modular architecture includes a clinical module for electrode labeling and localization, and a research module for automated data processing and electrode contact assignment. To ensure accessibility for users with limited programming and imaging expertise, we packaged iEEG-recon in a containerized format that allows integration into clinical workflows. We propose a cloud-based implementation of iEEG-recon and test our pipeline on data from 132 patients at two epilepsy centers using retrospective and prospective cohorts. RESULTS: We used iEEG-recon to accurately reconstruct electrodes in both electrocorticography and stereoelectroencephalography cases with a 30-min running time per case (including semiautomatic electrode labeling and reconstruction). iEEG-recon generates quality assurance reports and visualizations to support epilepsy surgery discussions. Reconstruction outputs from the clinical module were radiologically validated through pre- and postimplant T1-MRI visual inspections. We also found that our use of ANTsPyNet deep learning-based brain segmentation for electrode classification was consistent with the widely used FreeSurfer segmentations. SIGNIFICANCE: iEEG-recon is a robust pipeline for automating reconstruction of iEEG electrodes and implantable devices on brain MRI, promoting fast data analysis and integration into clinical workflows. iEEG-recon's accuracy, speed, and compatibility with cloud platforms make it a useful resource for epilepsy centers worldwide.

Resting-state background features demonstrate multidien cycles in long-term EEG device recordings.

BACKGROUND: Longitudinal EEG recorded by implanted devices is critical for understanding and managing epilepsy. Recent research reports patient-specific, multi-day cycles in device-detected epileptiform events that coincide with increased likelihood of clinical seizures. Understanding these cycles could elucidate mechanisms generating seizures and advance drug and neurostimulation therapies. OBJECTIVE/HYPOTHESIS: We hypothesize that seizure-correlated cycles are present in background neural activity, independent of interictal epileptiform spikes, and that neurostimulation may temporarily interrupt these cycles. METHODS: We analyzed regularly-recorded seizure-free data epochs from 20 patients implanted with a responsive neurostimulation (RNS) device for at least 1.5 years, to explore the relationship between cycles in device-detected interictal epileptiform activity (dIEA), clinician-validated interictal spikes, background EEG features, and neurostimulation. RESULTS: Background EEG features tracked the cycle phase of dIEA in all patients (AUC: 0.63 [0.56-0.67]) with a greater effect size compared to clinically annotated spike rate alone (AUC: 0.55 [0.53-0.61], p < 0.01). After accounting for circadian variation and spike rate, we observed significant population trends in elevated theta and beta band power and theta and alpha connectivity features at the cycle peaks (sign test, p < 0.05). In the period directly after stimulation we observe a decreased association between cycle phase and EEG features compared to background recordings (AUC: 0.58 [0.55-0.64]). CONCLUSIONS: Our findings suggest that seizure-correlated dIEA cycles are not solely due to epileptiform discharges but are associated with background measures of brain state; and that neurostimulation may temporarily interrupt these cycles. These results may help elucidate mechanisms underlying seizure generation, provide new biomarkers for seizure risk, and facilitate monitoring, treating, and managing epilepsy with implantable devices.

Disparities in seizure outcomes revealed by large language models.

Objective: Large-language models (LLMs) in healthcare have the potential to propagate existing biases or introduce new ones. For people with epilepsy, social determinants of health are associated with disparities in access to care, but their impact on seizure outcomes among those with access to specialty care remains unclear. Here we (1) evaluated our validated, epilepsy-specific LLM for intrinsic bias, and (2) used LLM-extracted seizure outcomes to test the hypothesis that different demographic groups have different seizure outcomes. Methods: First, we tested our LLM for intrinsic bias in the form of differential performance in demographic groups by race, ethnicity, sex, income, and health insurance in manually annotated notes. Next, we used LLM-classified seizure freedom at each office visit to test for outcome disparities in the same demographic groups, using univariable and multivariable analyses. Results: We analyzed 84,675 clinic visits from 25,612 patients seen at our epilepsy center 2005-2022. We found no differences in the accuracy, or positive or negative class balance of outcome classifications across demographic groups. Multivariable analysis indicated worse seizure outcomes for female patients (OR 1.33, p = 3×10-8), those with public insurance (OR 1.53, p = 2×10-13), and those from lower-income zip codes (OR ≥ 1.22, p ≤ 6.6×10-3). Black patients had worse outcomes than White patients in univariable but not multivariable analysis (OR 1.03, p = 0.66). Significance: We found no evidence that our LLM was intrinsically biased against any demographic group. Seizure freedom extracted by LLM revealed disparities in seizure outcomes across several demographic groups. These findings highlight the critical need to reduce disparities in the care of people with epilepsy.

Quantifying interictal intracranial EEG to predict focal epilepsy.

Introduction: Intracranial EEG (IEEG) is used for 2 main purposes, to determine: (1) if epileptic networks are amenable to focal treatment and (2) where to intervene. Currently these questions are answered qualitatively and sometimes differently across centers. There is a need for objective, standardized methods to guide surgical decision making and to enable large scale data analysis across centers and prospective clinical trials. Methods: We analyzed interictal data from 101 patients with drug resistant epilepsy who underwent presurgical evaluation with IEEG. We chose interictal data because of its potential to reduce the morbidity and cost associated with ictal recording. 65 patients had unifocal seizure onset on IEEG, and 36 were non-focal or multi-focal. We quantified the spatial dispersion of implanted electrodes and interictal IEEG abnormalities for each patient. We compared these measures against the "5 Sense Score (5SS)," a pre-implant estimate of the likelihood of focal seizure onset, and assessed their ability to predict the clinicians' choice of therapeutic intervention and the patient outcome. Results: The spatial dispersion of IEEG electrodes predicted network focality with precision similar to the 5SS (AUC = 0.67), indicating that electrode placement accurately reflected pre-implant information. A cross-validated model combining the 5SS and the spatial dispersion of interictal IEEG abnormalities significantly improved this prediction (AUC = 0.79; p<0.05). The combined model predicted ultimate treatment strategy (surgery vs. device) with an AUC of 0.81 and post-surgical outcome at 2 years with an AUC of 0.70. The 5SS, interictal IEEG, and electrode placement were not correlated and provided complementary information. Conclusions: Quantitative, interictal IEEG significantly improved upon pre-implant estimates of network focality and predicted treatment with precision approaching that of clinical experts. We present this study as an important step in building standardized, quantitative tools to guide epilepsy surgery.

iEEG-recon: A Fast and Scalable Pipeline for Accurate Reconstruction of Intracranial Electrodes and Implantable Devices.

Background: Collaboration between epilepsy centers is essential to integrate multimodal data for epilepsy research. Scalable tools for rapid and reproducible data analysis facilitate multicenter data integration and harmonization. Clinicians use intracranial EEG (iEEG) in conjunction with non-invasive brain imaging to identify epileptic networks and target therapy for drug-resistant epilepsy cases. Our goal was to promote ongoing and future collaboration by automating the process of "electrode reconstruction," which involves the labeling, registration, and assignment of iEEG electrode coordinates on neuroimaging. These tasks are still performed manually in many epilepsy centers. We developed a standalone, modular pipeline that performs electrode reconstruction. We demonstrate our tool's compatibility with clinical and research workflows and its scalability on cloud platforms. Methods: We created iEEG-recon, a scalable electrode reconstruction pipeline for semi-automatic iEEG annotation, rapid image registration, and electrode assignment on brain MRIs. Its modular architecture includes three modules: a clinical module for electrode labeling and localization, and a research module for automated data processing and electrode contact assignment. To ensure accessibility for users with limited programming and imaging expertise, we packaged iEEG-recon in a containerized format that allows integration into clinical workflows. We propose a cloud-based implementation of iEEG-recon, and test our pipeline on data from 132 patients at two epilepsy centers using retrospective and prospective cohorts. Results: We used iEEG-recon to accurately reconstruct electrodes in both electrocorticography (ECoG) and stereoelectroencephalography (SEEG) cases with a 10 minute running time per case, and ~20 min for semi-automatic electrode labeling. iEEG-recon generates quality assurance reports and visualizations to support epilepsy surgery discussions. Reconstruction outputs from the clinical module were radiologically validated through pre- and post-implant T1-MRI visual inspections. Our use of ANTsPyNet deep learning approach for brain segmentation and electrode classification was consistent with the widely used Freesurfer segmentation. Discussion: iEEG-recon is a valuable tool for automating reconstruction of iEEG electrodes and implantable devices on brain MRI, promoting efficient data analysis, and integration into clinical workflows. The tool's accuracy, speed, and compatibility with cloud platforms make it a useful resource for epilepsy centers worldwide. Comprehensive documentation is available at https://ieeg-recon.readthedocs.io/en/latest/.

Resting-state background features demonstrate multidien cycles in long-term EEG device recordings.

Background: Longitudinal EEG recorded by implanted devices is critical for understanding and managing epilepsy. Recent research reports patient-specific, multi-day cycles in device-detected epileptiform events that coincide with increased likelihood of clinical seizures. Understanding these cycles could elucidate mechanisms generating seizures and advance drug and neurostimulation therapies. Objective/Hypothesis: We hypothesize that seizure-correlated cycles are present in background neural activity, independent of interictal epileptiform spikes, and that neurostimulation may disrupt these cycles. Methods: We analyzed regularly-recorded seizure-free data epochs from 20 patients implanted with a responsive neurostimulation (RNS) device for at least 1.5 years, to explore the relationship between cycles in device-detected interictal epileptiform activity (dIEA), clinician-validated interictal spikes, background EEG features, and neurostimulation. Results: Background EEG features tracked the cycle phase of dIEA in all patients (AUC: 0.63 [0.56 - 0.67]) with a greater effect size compared to clinically annotated spike rate alone (AUC: 0.55 [0.53-0.61], p < 0.01). After accounting for circadian variation and spike rate, we observed significant population trends in elevated theta and beta band power and theta and alpha connectivity features at the cycle peaks (sign test, p < 0.05). In the period directly after stimulation we observe a decreased association between cycle phase and EEG features compared to background recordings (AUC: 0.58 [0.55-0.64]). Conclusions: Our findings suggest that seizure-correlated dIEA cycles are not solely due to epileptiform discharges but are associated with background measures of brain state; and that neurostimulation may disrupt these cycles. These results may help elucidate mechanisms underlying seizure generation, provide new biomarkers for seizure risk, and facilitate monitoring, treating, and managing epilepsy with implantable devices.

Improved Seizure Onset-Zone Lateralization in Temporal Lobe Epilepsy using 7T Resting-State fMRI: A Direct Comparison with 3T.

Objective: Resting-state functional magnetic resonance imaging (rs-fMRI) at ultra high-field strengths (≥7T) is known to provide superior signal-to-noise and statistical power than comparable acquisitions at lower field strengths. In this study, we aim to provide a direct comparison of the seizure onset-zone (SOZ) lateralizing ability of 7T rs-fMRI and 3T rs-fMRI. Methods: We investigated a cohort of 70 temporal lobe epilepsy (TLE) patients. A paired cohort of 19 patients had 3T and 7T rs-fMRI acquisitions for direct comparison between the two field strengths. Forty-three patients had only 3T, and 8 patients had only 7T rs-fMRI acquisitions. We quantified the functional connectivity between the hippocampus and other nodes within the default mode network (DMN) using seed-to-voxel connectivity, and measured how hippocampo-DMN connectivity could inform SOZ lateralization at 7T and 3T field strengths. Results: Differences between hippocampo-DMN connectivity ipsilateral and contralateral to the SOZ were significantly higher at 7T (pFDR=0.008) than at 3T (pFDR=0.80) when measured in the same subjects. We found that our ability to lateralize the SOZ, by distinguishing subjects with left TLE from subjects with right TLE, was superior at 7T (AUC = 0.97) than 3T (AUC = 0.68). Our findings were reproduced in extended cohorts of subjects scanned at either 3T or 7T. Our rs-fMRI findings at 7T, but not 3T, are consistent and highly correlated (Spearman Rho=0.65) with clinical FDG-PET lateralizing hypometabolism. Significance: We show superior SOZ lateralization in TLE patients when using 7T relative to 3T rs-fMRI, supporting the adoption of high-field strength functional imaging in the epilepsy presurgical evaluation.

DeepComBat: A Statistically Motivated, Hyperparameter-Robust, Deep Learning Approach to Harmonization of Neuroimaging Data.

Neuroimaging data from multiple batches (i.e. acquisition sites, scanner manufacturer, datasets, etc.) are increasingly necessary to gain new insights into the human brain. However, multi-batch data, as well as extracted radiomic features, exhibit pronounced technical artifacts across batches. These batch effects introduce confounding into the data and can obscure biological effects of interest, decreasing the generalizability and reproducibility of findings. This is especially true when multi-batch data is used alongside complex downstream analysis models, such as machine learning methods. Image harmonization methods seeking to remove these batch effects are important for mitigating these issues; however, significant multivariate batch effects remain in the data following harmonization by current state-of-the-art statistical and deep learning methods. We present DeepCombat, a deep learning harmonization method based on a conditional variational autoencoder architecture and the ComBat harmonization model. DeepCombat learns and removes subject-level batch effects by accounting for the multivariate relationships between features. Additionally, DeepComBat relaxes a number of strong assumptions commonly made by previous deep learning harmonization methods and is empirically robust across a wide range of hyperparameter choices. We apply this method to neuroimaging data from a large cognitive-aging cohort and find that DeepCombat outperforms existing methods, as assessed by a battery of machine learning methods, in removing scanner effects from cortical thickness measurements while preserving biological heterogeneity. Additionally, DeepComBat provides a new perspective for statistically-motivated deep learning harmonization methods.

Subcortical functional connectivity gradients in temporal lobe epilepsy.

BACKGROUND AND MOTIVATION: Functional gradients have been used to study differences in connectivity between healthy and diseased brain states, however this work has largely focused on the cortex. Because the subcortex plays a key role in seizure initiation in temporal lobe epilepsy (TLE), subcortical functional-connectivity gradients may help further elucidate differences between healthy brains and TLE, as well as differences between left (L)-TLE and right (R)-TLE. METHODS: In this work, we calculated subcortical functional-connectivity gradients (SFGs) from resting-state functional MRI (rs-fMRI) by measuring the similarity in connectivity profiles of subcortical voxels to cortical gray matter voxels. We performed this analysis in 24 R-TLE patients and 31 L-TLE patients (who were otherwise matched for age, gender, disease specific characteristics, and other clinical variables), and 16 controls. To measure differences in SFGs between L-TLE and R-TLE, we quantified deviations in the average functional gradient distributions, as well as their variance, across subcortical structures. RESULTS: We found an expansion, measured by increased variance, in the principal SFG of TLE relative to controls. When comparing the gradient across subcortical structures between L-TLE and R-TLE, we found that abnormalities in the ipsilateral hippocampal gradient distributions were significantly different between L-TLE and R-TLE. CONCLUSION: Our results suggest that expansion of the SFG is characteristic of TLE. Subcortical functional gradient differences exist between left and right TLE and are driven by connectivity changes in the hippocampus ipsilateral to the seizure onset zone.

Resting state functional connectivity demonstrates increased segregation in bilateral temporal lobe epilepsy.

OBJECTIVE: Temporal lobe epilepsy (TLE) is the most common type of focal epilepsy. An increasingly identified subset of patients with TLE consists of those who show bilaterally independent temporal lobe seizures. The purpose of this study was to leverage network neuroscience to better understand the interictal whole brain network of bilateral TLE (BiTLE). METHODS: In this study, using a multicenter resting state functional magnetic resonance imaging (rs-fMRI) data set, we constructed whole-brain functional networks of 19 patients with BiTLE, and compared them to those of 75 patients with unilateral TLE (UTLE). We quantified resting-state, whole-brain topological properties using metrics derived from network theory, including clustering coefficient, global efficiency, participation coefficient, and modularity. For each metric, we computed an average across all brain regions, and iterated this process across network densities. Curves of network density vs each network metric were compared between groups. Finally, we derived a combined metric, which we term the "integration-segregation axis," by combining whole-brain average clustering coefficient and global efficiency curves, and applying principal component analysis (PCA)-based dimensionality reduction. RESULTS: Compared to UTLE, BiTLE had decreased global efficiency (p = .031), and decreased whole brain average participation coefficient across a range of network densities (p = .019). Modularity maximization yielded a larger number of smaller communities in BiTLE than in UTLE (p = .020). Differences in network properties separate BiTLE and UTLE along the integration-segregation axis, with regions within the axis having a specificity of up to 0.87 for BiTLE. Along the integration-segregation axis, UTLE patients with poor surgical outcomes were distributed in the same regions as BiTLE, and network metrics confirmed similar patterns of increased segregation in both BiTLE and poor outcome UTLE. SIGNIFICANCE: Increased interictal whole-brain network segregation, as measured by rs-fMRI, is specific to BiTLE, as well as poor surgical outcome UTLE, and may assist in non-invasively identifying this patient population prior to intracranial electroencephalography or device implantation.

Quantifying trial-by-trial variability during cortico-cortical evoked potential mapping of epileptogenic tissue.

OBJECTIVE: Measuring cortico-cortical evoked potentials (CCEPs) is a promising tool for mapping epileptic networks, but it is not known how variability in brain state and stimulation technique might impact the use of CCEPs for epilepsy localization. We test the hypotheses that (1) CCEPs demonstrate systematic variability across trials and (2) CCEP amplitudes depend on the timing of stimulation with respect to endogenous, low-frequency oscillations. METHODS: We studied 11 patients who underwent CCEP mapping after stereo-electroencephalography electrode implantation for surgical evaluation of drug-resistant epilepsy. Evoked potentials were measured from all electrodes after each pulse of a 30 s, 1 Hz bipolar stimulation train. We quantified monotonic trends, phase dependence, and standard deviation (SD) of N1 (15-50 ms post-stimulation) and N2 (50-300 ms post-stimulation) amplitudes across the 30 stimulation trials for each patient. We used linear regression to quantify the relationship between measures of CCEP variability and the clinical seizure-onset zone (SOZ) or interictal spike rates. RESULTS: We found that N1 and N2 waveforms exhibited both positive and negative monotonic trends in amplitude across trials. SOZ electrodes and electrodes with high interictal spike rates had lower N1 and N2 amplitudes with higher SD across trials. Monotonic trends of N1 and N2 amplitude were more positive when stimulating from an area with higher interictal spike rate. We also found intermittent synchronization of trial-level N1 amplitude with low-frequency phase in the hippocampus, which did not localize the SOZ. SIGNIFICANCE: These findings suggest that standard approaches for CCEP mapping, which involve computing a trial-averaged response over a .2-1 Hz stimulation train, may be masking inter-trial variability that localizes to epileptogenic tissue. We also found that CCEP N1 amplitudes synchronize with ongoing low-frequency oscillations in the hippocampus. Further targeted experiments are needed to determine whether phase-locked stimulation could have a role in localizing epileptogenic tissue.

Subcortical Functional Connectivity Gradients in Temporal Lobe Epilepsy.

Background and Motivation: Functional gradients have been used to study differences in connectivity between healthy and diseased brain states, however this work has largely focused on the cortex. Because the subcortex plays a key role in seizure initiation in temporal lobe epilepsy (TLE), subcortical functional-connectivity gradients may help further elucidate differences between healthy brains and TLE, as well as differences between left (L)-TLE and right (R)-TLE. Methods: In this work, we calculated subcortical functional-connectivity gradients (SFGs) from resting-state functional MRI (rs-fMRI) by measuring the similarity in connectivity profiles of subcortical voxels to cortical gray matter voxels. We performed this analysis in 23 R-TLE patients and 32 L-TLE patients (who were otherwise matched for age, gender, disease specific characteristics, and other clinical variables), and 16 controls. To measure differences in SFGs between L-TLE and R-TLE, we quantified deviations in the average functional gradient distributions, as well as their variance, across subcortical structures. Results: We found an expansion, measured by increased variance, in the principal SFG of TLE relative to controls. When comparing the gradient across subcortical structures between L-TLE and R-TLE, we found that abnormalities in the ipsilateral hippocampal gradient distributions were significantly different between L-TLE and R-TLE. Conclusion: Our results suggest that expansion of the SFG is characteristic of TLE. Subcortical functional gradient differences exist between left and right TLE and are driven by connectivity changes in the hippocampus ipsilateral to the seizure onset zone.

Mapping hippocampal glutamate in mesial temporal lobe epilepsy with glutamate weighted CEST (GluCEST) imaging.

Temporal lobe epilepsy (TLE) is one of the most common subtypes of focal epilepsy, with mesial temporal sclerosis (MTS) being a common radiological and histopathological finding. Accurate identification of MTS during presurgical evaluation confers an increased chance of good surgical outcome. Here we propose the use of glutamate-weighted chemical exchange saturation transfer (GluCEST) magnetic resonance imaging (MRI) at 7 Tesla for mapping hippocampal glutamate distribution in epilepsy, allowing to differentiate lesional from non-lesional mesial TLE. We demonstrate that a directional asymmetry index, which quantifies the relative difference between GluCEST contrast in hippocampi ipsilateral and contralateral to the seizure onset zone, can differentiate between sclerotic and non-sclerotic hippocampi, even in instances where traditional presurgical MRI assessments did not provide evidence of sclerosis. Overall, our results suggest that hippocampal glutamate mapping through GluCEST imaging is a valuable addition to the presurgical epilepsy evaluation toolbox.

Interictal intracranial EEG asymmetry lateralizes temporal lobe epilepsy.

Patients with drug-resistant temporal lobe epilepsy often undergo intracranial EEG recording to capture multiple seizures in order to lateralize the seizure onset zone. This process is associated with morbidity and often ends in postoperative seizure recurrence. Abundant interictal (between-seizure) data is captured during this process, but these data currently play a small role in surgical planning. Our objective was to predict the laterality of the seizure onset zone using interictal (between-seizure) intracranial EEG data in patients with temporal lobe epilepsy. We performed a retrospective cohort study (single-center study for model development; two-center study for model validation). We studied patients with temporal lobe epilepsy undergoing intracranial EEG at the University of Pennsylvania (internal cohort) and the Medical University of South Carolina (external cohort) between 2015 and 2022. We developed a logistic regression model to predict seizure onset zone laterality using interictal EEG. We compared the concordance between the model-predicted seizure onset zone laterality and the side of surgery between patients with good and poor surgical outcomes. 47 patients (30 women; ages 20-69; 20 left-sided, 10 right-sided, and 17 bilateral seizure onsets) were analyzed for model development and internal validation. 19 patients (10 women; ages 23-73; 5 left-sided, 10 right-sided, 4 bilateral) were analyzed for external validation. The internal cohort cross-validated area under the curve for a model trained using spike rates was 0.83 for a model predicting left-sided seizure onset and 0.68 for a model predicting right-sided seizure onset. Balanced accuracies in the external cohort were 79.3% and 78.9% for the left- and right-sided predictions, respectively. The predicted concordance between the laterality of the seizure onset zone and the side of surgery was higher in patients with good surgical outcome. In conclusion, interictal EEG signatures are distinct across seizure onset zone lateralities. Left-sided seizure onsets are easier to distinguish than right-sided onsets. A model trained on spike rates accurately identifies patients with left-sided seizure onset zones and predicts surgical outcome.

Multi-contrast high-field quality image synthesis for portable low-field MRI using generative adversarial networks and paired data.

Introduction: Portable low-field strength (64mT) MRI scanners promise to increase access to neuroimaging for clinical and research purposes, however these devices produce lower quality images compared to high-field scanners. In this study, we developed and evaluated a deep learning architecture to generate high-field quality brain images from low-field inputs using a paired dataset of multiple sclerosis (MS) patients scanned at 64mT and 3T. Methods: A total of 49 MS patients were scanned on portable 64mT and standard 3T scanners at Penn (n=25) or the National Institutes of Health (NIH, n=24) with T1-weighted, T2-weighted and FLAIR acquisitions. Using this paired data, we developed a generative adversarial network (GAN) architecture for low- to high-field image translation (LowGAN). We then evaluated synthesized images with respect to image quality, brain morphometry, and white matter lesions. Results: Synthetic high-field images demonstrated visually superior quality compared to low-field inputs and significantly higher normalized cross-correlation (NCC) to actual high-field images for T1 (p=0.001) and FLAIR (p<0.001) contrasts. LowGAN generally outperformed the current state-of-the-art for low-field volumetrics. For example, thalamic, lateral ventricle, and total cortical volumes in LowGAN outputs did not differ significantly from 3T measurements. Synthetic outputs preserved MS lesions and captured a known inverse relationship between total lesion volume and thalamic volume. Conclusions: LowGAN generates synthetic high-field images with comparable visual and quantitative quality to actual high-field scans. Enhancing portable MRI image quality could add value and boost clinician confidence, enabling wider adoption of this technology.