Richard J. Staba

Diagnostic MRI evaluation of temporal lobe epilepsy (TLE) depends on the subjective visual interpretation of MRI images. These interpretations could be enhanced by quantitative artificial intelligence (AI) support tools. Humans often make sequential and conditional decisions during their radiological interpretations, such as whether an abnormality is present and if present, characterizing the abnormality. It is not known whether it is superior to train AI to treat every decision separately in a similar step-wise manner or to train a model holistically on all decisions simultaneously. Here, we analyzed three large epilepsy MRI datasets [n=3,676, 2,320 people with epilepsy and 1,356 healthy controls (HC)] to perform two tasks: 1) establish the presence of a TLE pattern on MRI and 2) determine TLE pattern lateralization. We compared Step-wise models that independently classify TLE versus HC and lateralize patients as left TLE (L-TLE) or right TLE (R-TLE), against a Simultaneous model trained to distinguish all three classes in a single step. To do this, 3D volumetric T1-weighted images were input into an EfficientNetV2 model multiple times to ensure reproducibility of results. Class prediction, model classification confidence, and saliency maps were output for interpretability. Step-wise models outperformed the Simultaneous model on both tasks (both ps<.001), with an average ∼2.8% accuracy increase for discriminating HC from TLE and an average 12.7% accuracy increase for distinguishing L-TLE from R-TLE. For both the Step-wise and Simultaneous models, important features discriminating TLE from HC included the known TLE limbic pattern involving the hippocampus, para-hippocampal cortical regions, cingulate cortex, and lateral temporal regions. However, there was less concordance between the Step-wise and Simultaneous models for the L-TLE versus R-TLE task (all Fisher’s Zs>10.5, ps<.001); the Step-wise model focused less on subcortical regions such as the thalamus and hippocampus and focused more on distributed cortical pathology. Across the two Step-wise models, 95.1% of TLE patients had accurate classifications in either HC versus TLE and/or L-TLE versus R-TLE tasks. These results included 69.6% of patients being both correctly labeled as TLE and lateralized, 13.9% being correctly labeled TLE but lateralized incorrectly, and 11.6% being lateralized correctly but not detected as TLE. These findings provide evidence that diagnostic tasks with simpler, Step-wise AI models may enhance diagnostic performance and interpretability in clinical workflows. Future AI clinical support tools can leverage this step-wise approach in the early identification of TLE-related structural patterns, supporting timely diagnosis and treatment decisions.

Objective: Tracing the propagation of high-frequency oscillations (HFOs) aids in localizing epileptogenic regions and improving surgical outcomes. We examined how hippocampal epileptogenicity influences the propagation properties of the HFOs it generates. Methods: We analyzed non-REM sleep stereo-EEG from 49 patients (68 hemispheres) with verified hippocampal contacts. Hippocampi were stratified by excitability: 28 seizure onset zone (SOZ), 22 more-irritative non-SOZ (>6 interictal epileptiform discharges [IED]/min), and 18 less-irritative non-SOZ (<6 IED/min). To isolate significant HFO propagation pathways, we constructed empirical temporal networks (maximum latency 150 ms) and validated them against 1,000 permutation-generated surrogates. We then compared the proportion of statistically significant propagating HFOs originating from hippocampal contacts across these groups. Results: We examined ripples on oscillation (RonO, 80-250 Hz) and fast ripples on oscillation (FRonO, 250-600 Hz). FRonO, but not RonO, rates were significantly elevated in hippocampal SOZ versus non-SOZ contacts (p<1e-9). Intrahippocampal RonO propagation proportion was highest in less-irritative non-SOZ compared to more-irritative non-SOZ (p<0.05) and SOZ (p<0.001). Across groups, we found no other differences in RonO or FRonO propagation proportions, including spread to other mesial-temporal structures or the neocortex. Significance: Intrahippocampal RonO propagation is proportionally greater in less-irritative non-SOZ tissue than in the epileptogenic hippocampus. Distinguishing physiological from pathological RonO using signal features alone remains challenging. Our work suggests these categories are not distinct; rather, RonO signals and their underlying hippocampal substrates likely exist on a continuous physiological-to-pathological spectrum. Furthermore, RonO propagation may serve as a novel metric to quantify hippocampal epileptogenicity. Key Points: Intrahippocampal propagation of ripples on oscillations (RonO, 80–250 Hz) is highest in the less-irritative hippocampus along the epileptogenicity spectrum despite their rates were similar. The proportions of hippocampal-to-mesial-temporal and hippocampal-to-neocortical RonO propagation remain constant across varying degrees of hippocampal epileptogenicity. Propagation proportions for fast RonO (fRonO, 250–600 Hz) do not differ significantly across the hippocampal epileptogenicity spectrum despite that their rates differ. Plain Language Summary: The hippocampus is a part of the brain that is crucial for memory. Sometimes, injured parts of the hippocampus create fast electrical brainwaves, called high-frequency oscillations (HFOs), which are linked to seizures. However, healthy parts of the hippocampus also create HFOs, and these are important for learning and memory. In this study, we wanted to know: do the brainwaves from an injured hippocampus spread differently through the brain than the brainwaves from a healthy one?

Objective: The development of posttraumatic epilepsy after traumatic brain injury (TBI) is potentially identifiable by measuring biomarkers of epileptogenesis, namely pathological high‐frequency oscillations (pHFOs). pHFOs are promising candidates, but it remains uncertain whether they can be detected early after TBI in clinical settings. This study was undertaken to determine the incidence and location of pHFOs as recorded from scalp and intracranial electroencephalography (EEG) during the first week after acute TBI and to determine the association of pHFOs with late posttraumatic seizures (PTS). Methods: We analyzed continuous EEG from 35 TBI patients with acute hemorrhagic TBI (Glasgow Coma Scale = 3–13) enrolled in the multicenter EpiBioS4Rx cohort. Automated pHFO detection was followed by independent experts' verification. The rate of two types of pHFO ripples (70–250 Hz) and fast ripples (250–500 Hz) were computed using scalp and intracranial EEG. Firth logistic regression models estimated associations between pHFOs rates and late PTS occurrence. Results: 16 of 35 patients (45.7%) developed late PTS. Verified scalp ripples were observed in 17 patients (48.6%), whereas no fast ripples were confirmed on expert review. Ripple activity was most frequent over frontal regions (13/17, 76.5%, p = .049, 95% confidence interval [CI] = .50–.93) and similarly frequent in perihemorrhagic locations (13/17, 76.5%, p = .049, 95% CI = .50–.93). Among 10 patients with intracranial EEG, verified ripples occurred in five (50%), including two of two with pericontusional strip electrodes. Significance: In critically ill patients with acute hemorrhagic TBI monitored in the intensive care unit, both scalp and intracranial EEG can detect pHFOs within the first week after injury, demonstrating the technical feasibility of capturing these signals in a real‐world clinical setting. Verified pHFOs were most frequently observed over frontal and perihemorrhagic regions, consistent with early perilesional hyperexcitability. pHFOs appear to be mechanistically grounded markers of perilesional hyperexcitability. Standardized, high‐sampling EEG studies with targeted pericontusional coverage are needed to establish prognostic performance for late PTS.

Introduction: Brain structural differences consistent with an older-appearing brain have been reported in people with epilepsy, but the extent to which these differences reflect clinical characteristics vs broader socioeconomic context is unclear. We investigated whether country-level socioeconomic factors are associated with neuroanatomical differences in adults with epilepsy using MRI-based age prediction, along with epilepsy subtype, sex, and clinical factors. Methods: Structural MRI and clinical data were collected from 26 epilepsy centres across 12 countries in the Americas, Australia, Europe, Asia and Africa. MRI-based age estimates were estimated using a previously developed prediction model trained on 29,175 healthy subjects. Brain predicted age difference (BrainPAD) was calculated as the difference between MRI-predicted brain age and chronological age. National gross domestic product (GDP) per capita and income inequality (Gini index) were obtained from the World Bank. Associations between BrainPAD and epilepsy subtype (temporal lobe epilepsy, extratemporal epilepsy, and genetic generalised epilepsy), national socioeconomic context (GDP per capita and Gini index), age and sex were assessed using regression models. Results: We analyzed 2,109 individuals with epilepsy and 1,041 healthy non-epilepsy controls (57% female; median age = 35; range 17-83). BrainPAD was higher in epilepsy than controls (β 4.2 years, SE 0.4; t=10.6), with increases ranging from 2.5 to 6 years across subtypes. Male sex was associated with 1 year higher BrainPAD relative to females (SE 0.33, t=3.12). There were no main effects of GDP or Gini index; however, significant interactions between were observed. The effect of epilepsy on BrainPAD was greater in countries with lower GDP per capita (t=-2.74) and higher income inequality (t=2.72). Conclusions: Clinical factors and socioeconomic context both influence brain structural ageing in epilepsy. These findings highlight the importance of geographic and economic diversity in neuroimaging research and underscore the relevance of global socioeconomic context when interpreting brain health measures.

High‐frequency oscillations (HFOs) were discovered more than 20 years ago, and since then they have been studied intensively in the context of epilepsy. HFOs encompass a broad spectrum of oscillations, typically ranging from 80 Hz to several kHz, that include both normal and pathological oscillations, documented in people with epilepsy and animal models. HFOs have drawn considerable attention due to their prominent roles in epileptogenesis, ictogenesis, and functional organization of epileptic tissue. We provide historical background on HFOs in epilepsy and summarize the current state of knowledge, synthesizing clinical and basic science content from the Third International Workshop on HFOs in Epilepsy. Over the years, the field has evolved from single‐center analysis of HFOs on invasive electroencephalographic recordings to recent multicenter studies and meta‐analysis, which have tempered the conviction or hope that HFOs are uniform, “one event fits all,” stand‐alone biomarkers. Instead, association of HFOs with other electrophysiological phenomena such as interictal spikes, seizures, and signal features like entropy have highlighted new ways to identify epileptogenic tissue. Advances in recording and analytical tools have significantly expanded their potential applications in both clinical and basic science settings. Several recent publications focus on how scalp HFOs can illuminate disease propensity, severity, and therapy responses. Moreover, it was recently discovered that HFOs are also present in experimental models of Alzheimer's disease, and research is ongoing regarding their relations to the HFOs found in epilepsy. Together, these developments highlight that HFOs represent an evolving research area, with significant inroads made over the years. Yet, key gaps in knowledge remain, and we propose five benchmark areas that warrant future research and advancement.

Thalamic neuromodulation is an emerging therapeutic strategy for patients in whom resection of the epileptogenic zone is not feasible. Yet the mechanisms by which thalamic stimulation modulates epileptogenic network dynamics remain unclear. We designed this study to test whether brief thalamic stimulation transiently modulates cortico-cortical activity and coupling within patient-specific epileptic networks. We studied 10 patients undergoing stereo-EEG with sampling of the anterior or centromedian thalamic nuclei. During wakefulness, we delivered low-frequency single pulse stimulations at 1 Hz and compared post-stimulation epochs (15–900 ms) with a pre-stimulation baseline. We quantified band-limited spectral power and Granger causality between bipolar cortical contacts, focusing on slow (3.0–13.9 Hz) and fast (14.0–79.9 Hz) activity. Contacts were grouped as seizure onset zone, early-propagating regions, or other regions, emphasizing the hemisphere ipsilateral to the stimulated nucleus. Thalamic stimulation produced a rapid, distributed modulation of cortical dynamics. Across patients, power in early-propagating regions decreased consistently in both slow and fast bands, whereas seizure onset zone and other regions showed smaller and less consistent changes, indicating reduced recruitment of cortical regions that often participate in seizure propagation. Directed interactions measured by Granger causality increased modestly at the network level, with the most consistent change between the seizure onset zone and other regions, and a significant increase in influence directed toward the seizure onset zone. These effects were observed with both nuclei despite heterogeneous implant strategies, suggesting a conserved, thalamus-initiated rebalancing of cortico-cortical coupling. Notably, when the anterior nucleus was stimulated in patients with limbic seizure onset zone, the power decrease in early-propagating regions and the increase in directed influence between the seizure onset zone and other regions were prominent, consistent with anterior thalamic projections to the limbic network. Taken together, these findings point to a rapid, nucleus-specific modulation that engages thalamic nodes while recruiting broader cortico-cortical influence toward the seizure onset zone. The sub-second markers—characterized by reduced power in early-propagating regions and selective increases in directed influence between seizure onset zone and other regions— support a model in which thalamic input modulates cortico-cortical coupling to limit propagation of epileptic activity from seizure onset zone to early-propagating regions. By characterizing network effects on the timescale of therapeutic stimulations, this work identifies mechanistic markers for network-informed neuromodulation, where transient EEG responses can guide nucleus selection and parameter titration. Prospective studies that link these markers to long-term seizure outcomes and stratify by thalamocortical anatomy are warranted.

Pathological high-frequency oscillations (HFOs 80-600 Hz) in intracranial EEG distinguish epileptogenic cortex. However, it is uncertain whether utilizing HFO measures for surgical planning improve epilepsy surgery seizure outcomes and minimize morbidity. The clinical gold standard for planning an epilepsy surgery involves consensus between epileptologists, radiologists, and neurosurgeons based on multimodality findings, and particularly the location of the seizure onset zone. We asked whether seizure freedom following epilepsy surgery could be accurately predicted using machine learning that uses measures of HFO features relative to the boundaries of a surgical resection or laser ablation. We detected and quantified HFOs from depth EEG contacts during 30-200 minutes of non-rapid eye movement sleep from 78 pre-surgical patients from three institutions. We trained a three-branch convolutional neural network (CNN) using 3 neuroanatomic features and 37 HFO derived features. The first and second CNN branches computed within and between patient differences, respectively, and the third branch contains the resected contacts that also influenced branches 1 and 2. We found that this HFO CNN labeled the seizure free patients with 92% accuracy using 5-fold cross-validation. These results suggest that a resection planned with the clinical gold standard can be prospectively evaluated by a HFO CNN approach to test whether the resection boundaries will achieve a seizure free outcome. Future work will explore utilizing the HFO CNN approach for counterfactual virtual resections constrained by a utility function to minimize morbidity.

Artificial intelligence (AI) is rapidly transforming our lives. Machine learning (ML) enables computers to learn from data and make decisions without explicit instructions. Deep learning (DL), a subset of ML, uses multiple layers of neural networks to recognize complex patterns in large datasets through end‐to‐end learning. DL has achieved remarkable success in challenging tasks such as facial recognition, self‐driving cars, and chatbots utilizing natural language processing. This progress extends to clinical practice and research in epilepsy. In this review, we provide an overview of AI applications in one of the most challenging yet crucial tasks in epilepsy: delineating the epileptogenic zone (EZ) using intracranial EEG. We discuss recent ML and DL real‐world workflows in EEG analysis, including preprocessing, training, evaluation, and interpretation. Then, we provide an overview of AI‐based approaches for EZ identification using interictal EEG, with a particular focus on high‐frequency oscillations. Lastly, we address the considerations and challenges that must be overcome to maximize the potential of AI‐based algorithms in EZ identification and biomarker discovery. We hope this review will foster discussions within the community on how to leverage AI to advance research and improve clinical practices in epilepsy.Plain Language Summary: AI has the potential to revolutionize epilepsy care by enhancing how clinicians pinpoint the brain regions responsible for seizures. This article highlights how AI can analyze interictal EEG (recorded between seizures) to detect subtle brain signals linked to epilepsy and to accurately define the epileptogenic zone. The AI‐based approach may lead to quicker diagnoses, more personalized treatments, and better surgical outcomes for patients with epilepsy.

High frequency oscillations (HFOs) are being incorporated into the presurgical evaluation of patients with epilepsy and also represent a rapidly evolving field in basic epilepsy research. Animal models have been pivotal in uncovering the mechanisms underlying HFOs and their role in epilepsy. In this review, we provide an overview of HFOs recorded in animal models, highlighting their relevance not only to epilepsy but also to neuropsychiatric and Alzheimer's diseases, suggesting a broader role in brain disorders. Then we explore recent advances, including innovative computational methods for the automated detection and analysis of HFOs. Building on this analysis, we propose guidelines for HFO identification and reporting with translational potential from animal models to humans. Finally, we discuss existing knowledge gaps and outline future directions for research in this rapidly evolving field.

Reproducibility of scientific data is a current concern throughout the neuroscience field. There are multiple on-going efforts to help resolve this problem. Within the preclinical neuroimaging field, the continued use of a region-of interest (ROI) type approaches combined with the well-known spatial heterogeneity of traumatic brain injury pathology is a barrier to the replicability and repeatability of data. Here we propose the conjoint use of an unbiased analysis of the whole brain after injury together with a population-based statistical analysis of sham-control brains as one approach that has been used in clinical research to help resolve this issue. The approach produces two volumes of pathology that are outside the normal range of sham brains, and can be interpreted as whole brain burden of injury. Using diffusion weighted imaging derived scalars from a tensor analysis of data acquired from adult, male rats at 2, 9 days, 1 and 5 months after lateral fluid percussion injury (LFPI) and in shams (n=73 and 12, respectively), we compared a data-driven, z-score mapping method to a whole brain and white matter-specific analysis, as well as an ROI-based analysis with brain regions preselected by virtue of their large group effect sizes. We show that the data-driven approach is statistically robust, providing the advantage of a large group effect size typical of a ROI analysis of mean scalar values derived from the tensor in regions of gross injury, but without the large multi-region statistical correction required for interrogating multiple brain areas, and without the potential bias inherent with using preselected ROIs. We show that the technique correctly captures the expected longitudinal time-course of the diffusion scalar volumes based on the spatial extent of the pathology and the known temporal changes in scalar values in the LFPI model.