Nishant Sinha

One-third of the world’s 70 million people with epilepsy have seizures uncontrolled by medication, and implantable devices are a promising treatment. These devices can detect impending seizures, missed medication and impaired cognition, but, like most medical devices, they cannot share this information with patients in real time. This is a missed opportunity for devices to learn from clinical events and teach patients how their behaviors affect their health. Here we present a device platform that converses with patients to help them manage their epilepsy. The inpatient prototype links scalp and intracranial electroencephalography (EEG) to a secure, large language model-based artificial intelligence (AI) that communicates bidirectionally through a smartphone app. The AI processes biomarkers of sleep, medication, cognition and seizure probability, notifies patients of clinical events and trends, records symptoms and responses, and uses their feedback to retrain its seizure detector, reducing false alarms. We validated this platform in 13 patients who are undergoing video-EEG monitoring for epilepsy (scalp, n = 11; intracranial, n = 2). In a pseudoprospective simulation, seizure detectors improved precision from patient feedback alone, without expert intervention. We propose this system as a scalable, safeguarded platform for medical devices that can rapidly adapt to patient and provider needs across epilepsy and other medical conditions.

Objective: Cortico-cortical evoked potentials (CCEPs) are used during intracranial EEG (iEEG) monitoring to probe effective connectivity and guide epilepsy surgery. However, the factors governing spatial distribution and variability of stimulation-evoked responses across patients remain incompletely understood, making it difficult to distinguish physiological variation from pathological responses. We aimed to develop interpretable models of CCEP responses that integrate normative and patient-specific features, and to assess how these relationships are influenced by distance, brain-state, and epileptogenicity. Methods: We analyzed single-pulse stimulation data from 34 patients with drug-resistant epilepsy undergoing stereo-EEG monitoring. Significant CCEP responses were defined using N1 peak amplitude. For each stimulation-response pair, we derived multimodal features capturing Euclidean distance, effective connectivity priors from F-TRACT, a published population atlas of CCEP response probabilities, normative structural connectivity and communication metrics, and patient-specific functional connectivity estimated from iEEG during baseline and pre-stimulation periods. Lasso-regularized logistic regression models were trained to predict response using leave-one-patient-out cross-validation, and feature importance was examined across spatial scales and seizure onset zone (SOZ) involvement. Results: Performance to predict stimulation-evoked responses improved with the integration of multiple feature domains, with the full model achieving the highest accuracy (median AUC = 0.84). Feature relevance varied with spatial scale: short-range responses were dominated by distance and normative priors, whereas long-range responses depended strongly on patient-specific functional connectivity, especially when measured from pre-stimulation periods. Prediction errors were significantly higher for stimulation-response pairs within the SOZ, indicating deviations from normative connectivity-response relationships in epileptogenic networks. Significance: CCEP responses reflect an interaction between spatial constraints, network architecture, and patient-specific brain state. Multimodal modeling of CCEPs provides a principled framework to interpret stimulation responses and may offer complementary markers for identifying epileptogenic networks during presurgical evaluation.

Background: In glioblastoma, tumor-related epilepsy (TRE) is paradoxically associated with longer survival, but the mechanism is unclear. We hypothesized that structural connectivity is associated with seizure susceptibility and survival, and tested whether these associations are independent or mediated. Methods: We retrospectively studied glioblastoma patients with preoperative diffusion-weighted MRI. Fractional anisotropy-weighted connectivity matrices were constructed using lesion-aware tractography. Mean edge weight and local efficiency were computed per Yeo 7-network and with global efficiency at hemisphere and whole-brain levels. Associations with TRE, overall survival (OS), and progression-free survival (PFS) were assessed using logistic regression and Cox proportional hazards models with false discovery rate correction. Causal mediation analysis tested whether connectivity and TRE mediate each other’s relationship with overall survival. Results: We included 420 glioblastoma patients. Of 406 classifiable for TRE, 217 (53%) had TRE. TRE was associated with higher local efficiency across all seven contralateral networks ( d= 0.42-0.49, q< 0.010) and five of seven ipsilateral networks ( d =0.44-0.49, q< 0.050). In adjusted models, higher contralateral global efficiency was associated with longer OS (HR 0.84, q= 0.021), with no metric associated with PFS. TRE did not mediate the connectivity-OS relationship. Contralateral local efficiency partially mediated the TRE-OS association when postoperative KPS was excluded from the model (ACME 16.6 days, 7.8% mediated, p =0.029), though this was attenuated when KPS was included ( p =0.074). Conclusions: Structural connectivity may partially mediate the seizure-survival association, suggesting preserved network architecture may contribute to longer survival in glioblastoma patients with TRE, though most of the effect operates through other pathways. Key Points: Tumor-related epilepsy in glioblastoma was associated with relatively greater structural connectivity. Higher contralateral global efficiency related to longer overall survival. Contralateral local efficiency may partly mediate the seizure-survival link. Importance of the Study: Seizures in glioblastoma are paradoxically associated with longer survival, yet the biological basis of this relationship has remained unknown. Using lesion-aware diffusion tractography in 420 patients, we found that the same structural network measures relate to both seizure susceptibility and survival. Patients with tumor-related epilepsy had relatively greater white matter connectivity, and higher contralateral connectivity was independently associated with longer overall survival. Causal mediation analysis showed that connectivity accounted for a small fraction of the seizure-survival association, suggesting that seizures mark, rather than confer, the survival advantage, and that both may reflect a shared upstream driver not captured here. These findings identify structural connectivity as a shared correlate and candidate imaging biomarker for seizure risk and prognosis in glioblastoma.

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: Interictal spikes have been proposed as a biomarker for both localizing seizure onset zones (SOZ) and tracking changes in seizure risk with neurostimulation in patients with drug-resistant epilepsy. Electrical stimulation can modulate spike rates acutely, and it has been proposed that measuring this modulation can help localize the SOZ. However, it is unclear whether stimulation-induced spike rate changes reflect epilepsy-specific pathology in the stimulated network or simply intrinsic regional excitability, which limits our understanding of their utility in epilepsy surgery planning. Methods: We analyzed low-frequency stimulation (LFS; 1 Hz) applied during a clinical seizure-induction protocol systematically targeting multiple brain regions in 43 patients with drug-resistant epilepsy undergoing intracranial EEG monitoring. A validated, automated spike detector was used to quantify pre-, during-, and post-stimulation spike rates. We tested whether the stimulation-evoked spike rate response (i) tracks the expected change in seizure risk from a seizure induction protocol, (ii) varies with anatomical stimulation site and epilepsy localization, (iii) localizes the SOZ beyond baseline spike rate, and (iv) is accompanied by changes in spike morphology. Results: Nearby LFS acutely increased spike rates in high-spiking channels (inter-stimulation median 2.25 vs. during-stimulation 4.25 spikes/min; p < 0.001), with effects attenuating with distance and resolving within approximately 30 seconds of stimulation offset. Mesial temporal lobe stimulation produced the largest increase in nearby spike rates relative to temporal neocortex and other cortex (Kruskal-Wallis p = 0.003), but this effect did not differ between patients with and without mesial temporal lobe epilepsy. A random forest classifier incorporating stimulation-evoked modulation features achieved an AUC of 0.787, comparable to a resting-state spike model (AUC 0.747; DeLong p = 0.81), indicating that stimulation-evoked spike changes do not add localizing information beyond resting-state spike rates. Stimulation produced a small but significant shift in spike morphology toward broader, higher-amplitude discharges (PERMANOVA p < 0.001), consistent with recruitment of a broader neuronal population. Significance: LFS-evoked increases in interictal spike rates reflect intrinsic regional excitability, greatest in the mesial temporal lobe, rather than epilepsy-specific pathology, and do not improve SOZ localization over resting-state spike rates. These results argue against using the change in spikes with stimulation to localize the SOZ. On the other hand, the transient spike rate increase induced by a pro-epileptic protocol supports the acute change in spike rate as a biomarker of the effect of stimulation on seizure risk, with potential to guide parameter selection for epilepsy neuromodulation.

About half of patients who undergo epilepsy surgery for drug-resistant epilepsy have seizure recurrence, supporting the need for approaches that more accurately identify the epileptogenic zone , defined as the brain areas whose removal causes cessation of seizures. Altered network connectivity has emerged as a candidate biomarker of the epileptogenic zone, but how connectivity is altered in the epileptogenic zone remains uncertain, with prior studies reporting inconsistent results. We hypothesized that a difference in intrinsic versus extrinsic connectivity of the epileptogenic zone may explain prior discrepant findings. We studied a multicenter cohort of adult and pediatric patients who underwent intracranial EEG recording and brain stimulation as part of epilepsy surgery planning. We measured spontaneous connectivity using Pearson correlation and perturbational connectivity using stimulation evoked potentials, modeling the connectivity according to the location of contacts in relation to the seizure onset zone (SOZ) while controlling for inter-electrode distance. We analyzed 79 patients (37 adults, 42 children). For both adult and pediatric patients, resting connectivity was higher within compared to outside the SOZ, but resting connectivity between SOZ and non-SOZ contacts was reduced. Stimulation connectivity followed a similar pattern, with elevated within-SOZ connectivity but reduced connectivity between SOZ and non-SOZ. The results support the hypothesis that the epileptogenic zone is disconnected from the rest of the brain but intrinsically hyperconnected. This result helps reconcile prior inconsistencies across studies, aligns with the results of basic science studies, and suggests that future translational work should model this heterogeneous pattern to increase the yield of using connectivity to localize the epileptogenic zone.

Intracranial electroencephalographic (iEEG) connectivity analysis is a promising method to localize epileptic networks and guide surgical planning in focal drug‐resistant epilepsy. Despite numerous studies exploring its utility, the added value of iEEG connectivity over standard clinical presurgical evaluation remains unclear. We assess the current evidence on the efficacy of iEEG connectivity analyses to improve seizure outcomes following epilepsy surgery through a systematic review and meta‐analysis. Following PRISMA (Preferred Reporting Items for Systematic Reviews and Meta‐Analyses) reporting guidelines, we searched PubMed and Embase for studies (2006–2024) of adult focal drug‐resistant epilepsy patients who underwent surgical resection or ablation, reported outcomes at least 1 year postsurgery, and used iEEG connectivity analysis to localize networks. Reviews, nonhuman studies, and studies lacking iEEG connectivity analysis or network localization were excluded. We derived classification metrics (true/false positives/negatives) based on concordance between iEEG findings, clinical localization, and outcome. Subgroup meta‐analyses and meta‐regressions determined differences by seizure type, lesion status, and analysis approach. Of 2881 studies screened, 25 met criteria ( n = 909). The pooled odds ratio comparing seizure outcome prediction using iEEG connectivity versus standard clinical evaluation was 1.36 (95% confidence interval = 1.10–1.69, p = .004), indicating a significant overall benefit. Subgroup analyses found no significant differences by directionality, modeling method (linear/nonlinear), or iEEG epoch (interictal/peri‐ictal). Meta‐regression revealed greater added value of iEEG connectivity in studies with higher proportions of non‐seizure‐free patients following surgery for temporal lobe or lesional epilepsy. However, no individual study achieved statistical significance on its own, reflecting limited power and lack of individual patient‐level data. Power analysis confirmed that detecting a clinically meaningful effect requires substantially larger, potentially multicenter datasets. iEEG connectivity analysis offers modest but consistent increased value over standard clinical methods to predict seizure freedom in adult patients with focal drug‐resistant epilepsy. For clinical translation, we propose recommendations for future studies to address sample size limitations, standardize reporting, and prioritize individual patient‐level data sharing.

Annotating seizure onset and spread in intracranial EEG is essential for epilepsy surgical planning, yet manual annotation is unreliable and cannot scale to large datasets. We introduce Neural Dynamic Divergence (NDD), an unsupervised framework that detects seizure activity by measuring deviation from patient-specific baseline neural dynamics using autoregressive models. NDD requires no labeled training data and adapts to individual patients, channels, and brain states. Validating against expert consensus annotations from 46 seizures, NDD achieves human-level agreement ( ϕ = 0.58 vs. inter-rater ϕ = 0.64) and outperforms existing algorithms on 1,019 seizures with soft labels (AUROC = 0.87). We demonstrate clinical utility by automatically annotating 2,017 seizures, revealing that seizure spread patterns distinguish epilepsy subtypes and predict surgical outcomes. NDD also generalizes to continuous ICU scalp EEG monitoring (AUROC = 0.77). We provide NDD as an open-source Python package to enable scalable seizure annotation across research centers.

One-third of the world’s 70 million people with epilepsy have seizures that are not controlled by medication; and implantable devices are an exciting option for treatment. These devices improve seizure control and can detect impending attacks, missed medication, and impaired cognition. Unfortunately, they have no way to share this information with their hosts in real-time – a limitation common to most medical devices. This is a missed opportunity for implants and wearables to learn from patients, focus on what matters most to them, and teach them how their behavior affects their health. Here, we present a device platform that converses with patients and learns to co-manage epilepsy. The inpatient pro-totype links scalp and intracranial EEG (electroencephalograms) to secure large language models that communicate freely and bidirectionally with their hosts through a smartphone app. An AI agent ingests biomarkers of sleep, medication level, cognition, and seizure risk extracted from brain activity. It con-verses with patients to inform them of clinical events and physiological trends, records their symptoms, responses, and behaviors, and automatically retrains itself to improve performance. Both patients and the AI agent can initiate conversations to teach each other and personalize interactions. We demon-strate this platform in 13 patients undergoing inpatient video-EEG monitoring for epilepsy and validate its performance. Algorithms for detecting seizures optimized their precision over several days without expert intervention – in contrast to the months of iterative, in-person physician programming currently required. Patients responded positively to messages regarding sleep, cognition, and seizure risk while rating the system as highly usable. The platform includes several safeguards, including a system for further algorithm fine-tuning using efficient expert review, and features that ensure data security and regulate communication content. Further work will link other biosensors to measure behavior, improve performance, and optimize therapeutic stimulation. We propose this system as a scalable platform for medical devices that can rapidly adapt to patient and provider needs; one that is broadly adaptable to improving care for many medical conditions.

Automated seizure detection and localization from intracranial EEG requires validated benchmark datasets with expert annotations, yet existing open datasets lack multi-expert consensus annotations and exclude stimulation-induced seizures. We present stereotactic EEG recordings from 83 seizures (46 spontaneous, 37 stimulation-induced) across 32 patients (19 from the University of Pennsylvania, 13 from the Children’s Hospital of Philadelphia) with drug-resistant epilepsy. Three board-certified epileptologists independently annotated each seizure for onset time, onset channels, and channels seizing at 10 seconds post-onset using a standardized protocol. Consensus annotations were determined through majority voting. Inter-rater agreement was κ = 0.64 for onset channels and κ = 0.62 for spread channels. Individual rater agreement with consensus was κ = 0.81 for onset and κ = 0.80 for spread. Agreement metrics did not differ between spontaneous and stimulation-induced seizures. All data follow Brain Imaging Data Structure (BIDS) standards and include electrode localizations, patient demographics, and clinical outcomes. This dataset enables the validation of seizure onset and spread detection and localization against human expert performance and supports comparative analysis of seizure networks across spontaneous and stimulation-induced seizures.