Analysis of EEG univariate features for epileptic seizures.

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 prototype 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 awareness, sleep, medication level, cognition, and seizure risk extracted from brain activity. It converses 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 demonstrate 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.

Background/Objectives: Surgical resection of the epileptogenic zone (EZ) is the most effective treatment for drug-resistant epilepsy, yet its success relies on accurate localization. Interictal spikes and high-frequency oscillations (HFOs) are important electrophysiological biomarkers for the EZ, but their visual identification in long-term recordings is labor-intensive, subjective, and prone to inter-rater variability. Artificial Intelligence (AI) offers a promising solution for automating this process. This review provides a comprehensive synthesis of AI-based methods for detecting both spikes and HFOs across invasive intracranial electroencephalography (iEEG) and non-invasive scalp EEG. Methods: We organized this review by recording modality (iEEG, scalp EEG) and biomarker type (spikes, HFOs), tracing the methodological evolution of biomarker detection, ranging from early rule-based heuristics to feature-engineered machine learning (ML) and modern end-to-end deep learning (DL). We then described the clinical application of these detected events, specifically their role in localizing the epileptogenic zone and defining state-dependent prognostic markers. Next, we provided an overview of validated software tools and annotated datasets essential for benchmarking new algorithms. Finally, we listed critical challenges and offered perspective on the future directions of this field. Results: In iEEG, spike and HFO detection methods have matured from prone-to-error energy thresholds to advanced deep learning and neuromorphic systems that significantly reduce false positives and enable real-time network analysis. In scalp EEG, detection analysis has progressed from manual heuristics to robust deep neural networks that successfully overcome low signal-to-noise ratios and muscle artifacts to achieve expert-level diagnostic performance. The clinical utility of AI-based detection includes precise epileptogenic zone localization, specifically through spike-HFO co-occurrence, and the monitoring of state-dependent fluctuations for objective prognostication. To ensure reliable deployment, researchers need to validate models against standardized datasets and commercial benchmarks while tailoring architectures to specific data scales and real-time processing constraints. Conclusions: AI-driven automation is essential for realizing the clinical potential of EEG biomarkers, yet widespread adoption is currently impeded by several factors, such as the lack of interpretability in “black box” models, the generalization gap across diverse centers, computational constraints preventing real-time implementation, labeling bottlenecks, and complex regulatory and ethical requirements. Future research should prioritize Explainable AI (XAI), multi-center validation, self-supervised learning, and the development of “human-in-the-loop” workflows to robustly integrate automated detection into routine presurgical evaluation.

Sleep architecture, including spindles and slow oscillations, is disrupted in Alzheimer disease (AD). How changes in these sleep elements relate to cognitive decline is less clear. Our objectives were to examine changes in sleep macroarchitecture and microarchitecture in early clinical stages of AD compared with older adult controls (CTLs) and to investigate their associations with longitudinal cognitive change. This was both a cross-sectional and longitudinal study performed at Mass General Brigham Hospitals, where participants with early AD or CTLs underwent overnight ambulatory scalp EEG and longitudinal cognitive testing. We compared sleep microarchitectural features extracted from the EEG, including spindle activity, across the brain topography and between groups. We then performed longitudinal analyses using mixed-effects models to test the association of these sleep features with changes in cognition on the Montreal Cognitive Assessment (MoCA), collected annually for up to 7 years. AD (n = 47, mean age 74.1 years, 66% female) and CTL (n = 43, mean age 72.6 years, 56% female) groups spent a similar proportion of sleep time in each stage of sleep. Sleep efficiency, however, was lower in the AD group (mean: CTL 75.1% vs AD 70.9%; p = 0.034). We found a significant reduction in spindle range power (11-16 Hz) in patients with AD compared with CTLs, particularly in the temporal regions (mean normalized power at EEG channels T3/T4: CTL 3.13 ± 1.13 vs AD 2.48 ± 1.01; p = 0.005). In participants with longitudinal MoCA scores (AD = 26, CTL = 25), reduced temporal lobe spindle density (β = 0.61, 95% CI 0.35-0.87; false discovery rate [FDR]-adjusted p < 0.001) and temporal lobe spindle power (β = 0.56, 95% CI 0.22-0.88; FDR-adjusted p = 0.005) were each associated with a faster rate of cognitive decline. Temporal lobe sleep spindle activity is reduced in early clinical stages of AD and is associated with a faster rate of cognitive decline. Our results underscore the importance of including temporal lobe measurements when assessing sleep neurophysiology in AD, which is not standard in polysomnography. Future work examining the relationship between AD biomarkers and reduced spindle activity is needed to elucidate the potential mechanisms underlying these findings.

A causal attention network with time frequency channel feature fusion for epileptic seizure prediction.

Performance evaluation of the earphone-type electroencephalogram for alpha wave detection.

Usefulness of the Scalp Electroencephalography over Bone Defect for the Decoding of Muscle Activity.

Epilepsy diagnosis and treatment monitoring are hindered by the episodic, heterogeneous expression of seizures and by normal-appearing scalp EEG in many patients. We previously described paroxysmal slow-wave events (PSWEs)-brief epochs of broadband slowing detectable on EEG. Here, using intracerebral and epidural recordings in a paraoxon rat model of temporal lobe epilepsy, we show that PSWEs arise preferentially in temporo-frontal networks, co-occur with global slowing, and increase during both spontaneous and pharmacologically induced seizures. Epidurally recorded PSWEs were temporally coupled to deep temporal discharges and were bidirectionally modulated by GABAergic agents (increased with pentylenetetrazol and decreased with pentobarbital). In long-term video-EEG monitoring (LTM) patients with temporal lobe epilepsy, simultaneous stereo-EEG and scalp EEG showed that scalp PSWEs mirrored hippocampal spike-and-wave activity and were more often observed in the preictal and ictal periods than during interictal baseline. These data indicate that surface PSWEs can index remote epileptiform activity and support their use as a quantitative, noninvasive biomarker for detecting EEG-silent deep foci and for pharmacodynamic.

Wearable and implantable microelectrodes are widely used to investigate brain function and treat neurological disorders. Here, we present a dual-function electrophysiological and electrochemical (EC) microelectrode composed of Cu-doped graphene nanosheets. The graphene nanosheets were vertically grown via chemical vapor deposition (CVD) and possess excellent electrical properties while exhibiting enhanced electrochemical (EC) activity due to the in situ-doped Cu. An electroencephalography (EEG) headband for humans, adrenaline (Adr) measurement device, and indwelling needle cores were successfully assembled using these graphene microelectrodes. The results of in vivo rat experiments revealed that the Cu-vertical graphene (VG) microelectrodes can accurately record intracranial electrocorticogram signals. Moreover, the EC signals could be used to test Adr levels at concentrations ranging from 0.5 to 800 μM at a pH range of 5-9.0, with detection limits within 0.024-0.112 μM. In human trials, the microelectrodes can recognize differences between EEG signals arising from the left and right frontal lobes and quickly respond to Adr levels in bodily fluids. Furthermore, when the left or right arms of the volunteers were raised, the scalp EEG signals for the left and right frontal lobes could be differentiated. Therefore, Cu-VG microelectrodes are well suited for wearable and portable human-machine interface sensors for rapid monitoring of brain function and neurotransmitter levels in clinical and point-of-care settings. This study presents a reliable human-machine interface sensor that can be utilized to correlate EEG signals and neurotransmitter regulation.

This study aims to establish a method-integrative framework for emotion-oriented architectural image generation. The framework combines Stable Diffusion with targeted LoRA (Low-Rank Adaptation), a lightweight and parameter-efficient fine-tuning approach, together with ControlNet-based structural constraints, to examine how controllable design-element manipulations influence emotional responses. The methodology follows a closed-loop “generation–evaluation” workflow, with each LoRA module independently targeting a single design element. Guided by the relaxation–arousal emotional dimension, the framework is evaluated using subjective ratings and electroencephalogram (EEG) measures. Twenty-seven participants viewed six architectural space categories, each comprising four conditions (baseline, color, material, and form modification). EEG α/β power ratio (RAB) served as the primary neurophysiological marker of arousal. Statistical analysis indicated that LoRA-based modifications of design elements produced distinct emotional responses: color and material changes induced lower arousal, whereas changes in form elicited a bidirectional pattern involving relaxation and arousal. The right parietal P4 electrode site showed the most sensitive emotional response to design element changes, with consistent statistical significance. P4 is a human scalp EEG location associated with cortical activity related to visuospatial processing. Descriptive results suggested opposite directional effects with similar intensity trends; however, linear mixed-effects model (LMM) inference did not support significant group-level linear coupling, indicating individual variation. This study demonstrates the feasibility of emotion-guided architectural image generation, showing that controlled manipulation of color, material, and form can elicit measurable emotional responses in human brain activity. The findings provide a methodological basis for future multimodal, adaptive generative systems and offer a quantitative pathway for investigating the relationship between emotional states and architectural design elements.

Background/Objectives: Precise identification of seizure onset zones (SOZs) and their propagation pathways is essential for effective epilepsy surgery and other interventional therapies and is typically achieved through invasive electrophysiological recordings such as intracranial electroencephalography (EEG). Previous research has demonstrated that analyzing information flow patterns, particularly in high-frequency oscillations (>80 Hz) using parametric and Wilson algorithm (WL)-based nonparametric Granger causality (GC), is valuable for SOZ identification. In this study, we analyzed scalp EEG recordings from epilepsy patients using an alternative nonparametric GC approach based on spectral density matrix factorization via the Janashia–Lagvilava algorithm (JLA). The aim of this study is to evaluate the effectiveness of JLA-based matrix factorization in nonparametric GC for noninvasively identifying seizure onset zones from ictal EEG recordings in patients with drug-resistant epilepsy. Methods: Two regions of interest (ROIs) in pairs were isolated across different time epochs in six patients referred for presurgical evaluation. To apply the nonparametric Granger causality (GC) estimation approach to the EEG recordings from these regions, the cross-power spectral density matrix was first computed using the multitaper method and subsequently factorized using the JLA. This factorization yielded the transfer function and noise covariance matrix required for GC estimation. GC values were then obtained at different prediction time steps (measured in milliseconds). These estimates were used to confirm the visually suspected seizure onset regions and their propagation pathways. Results: JLA-based spectral factorization applied within the Granger causality framework successfully identified SOZs and their propagation patterns from scalp EEG recordings, demonstrating alignment with positive surgical outcomes (Engel Class I) in all six cases. Conclusions: JLA-based spectral factorization in nonparametric Granger causality shows strong potential not only for accurate SOZ localization to support diagnosis and treatment, but also for broader applications in uncovering information flow patterns in neuroimaging and computational neuroscience.

The purpose of this study was to assess the change in frequency and distribution of focal interictal epileptiform discharges (IEDs) as measured on scalp EEG after anti-seizure medications (ASMs) were weaned in the epilepsy monitoring unit. We retrospectively reviewed the EEG of patients with focal epilepsy on a single ASM. A two-hour EEG epoch was selected at sleep onset during the first day of admission and defined as the high-ASM epoch. This was compared to a two-hour low-ASM epoch at sleep onset after the ASM was weaned, at least 6 h before or after a seizure. IEDs were manually counted and characterized. A total of 115 patients were included. For those on levetiracetam, there was a significant increase in IED quantity when comparing the high-ASM to the low-ASM epoch (mean 40.6 to 71.4, p < 0.001). For those on sodium channel blockers, there was a non-significant trend towards a decrease in IED quantity as the ASM dose was decreased (p = 0.065). There was no statistically significant change found for other individual ASMs. For the cohort, 12 patients had IED observed only on the low-ASM epoch (which were not present on the high-ASM epoch), 6 of which were treated with levetiracetam. In summary, our findings showed weaning of levetiracetam was associated with a significant increase in IEDs whereas other ASMs were not. Some populations of IEDs were only seen after ASMs were weaned. These findings suggest that different ASMs may have unique effects on IEDs when weaned.

Objective: Patient self-report is known to be an inaccurate reflection of true seizure frequency in persons with epilepsy. The current study aimed to assess the safety and performance of the Minder system, a bilateral subscalp electroencephalographic (EEG) acquisition system for continuous long-term EEG recording. Methods: This prospective, multicenter first-in-human study enrolled adult patients with focal or generalized epilepsy and at least 2 seizures per month. The primary outcome was adverse events (AEs) in the first 6 months of implantation. Secondary analyses determined whether normal neurophysiological signals, interictal discharges, and seizures seen on scalp video EEG monitoring were identifiable on subscalp recordings, and signals were rated for clarity on subscalp and 2-channel scalp EEG recordings (1 = not recognizable, 5 = clear). Subscalp data were reviewed in relation to events reported in 6-month seizure diaries. Results: Twenty-six subjects were implanted between November 2019 and July 2023. No serious device- or implant procedure-related AEs were reported. The most common device-related AEs were mild or moderate postsurgical pain, headache, or scalp pain/paresthesia (9/26, 35%). All sleep spindles, chewing artifacts, interictal discharges, and electrographic seizures observed on scalp recordings (25 seizures from 8 patients) were identified on subscalp recordings and given higher clarity ratings compared to 2-channel scalp recordings (median seizure clarity rating was 3 for both scalp and subscalp EEG, range = 1-5, P = .0025). Subscalp recordings captured seizures from diverse seizure focus locations, including frontal and mesial temporal seizure foci and hypothalamic hamartoma. Bilateral recording revealed clinically relevant findings not possible with unilateral recordings (6/26 patients, 23%). Findings of potential clinical utility were identified on manual review of 6-month recordings in most patients (23/26, 88%). Significance: This study demonstrates the safety and performance of the Minder bilateral subscalp EEG acquisition system for long-term seizure monitoring in patients with epilepsy. Bilateral hemisphere coverage captured seizures in a diverse patient group and permitted lateralization of events.

Epilepsy is a chronic neurological disorder that significantly affects the quality of life (QoL), often causing irreversible brain damage and physical impairment. Electroencephalography (EEG) signal analysis is crucial for monitoring epilepsy, enabling early seizure detection and timely intervention. Effective seizure detection requires the identification of interpretable features from the EEG signal to improve clinical outcomes. This study proposes a novel interpretable multi-view feature learning approach (IMV-FL), in which the time-domain signals and Discrete Fourier Transform (DFT) are applied to convert the time-domain EEG signal into frequency-domain representations. To develop initial multiview feature extraction and compression, spatial and temporal morphological features are extracted from optimal layers of ResNet and Long Short-Term Memory (LSTM) models, with feature compression performed using a Deep Neural Network (DNN). To construct an interpretable multi-view feature fusion, linear and nonlinear properties are calculated for the feature and with fusion strategies. The selected features are processed using the Mutual Information-Based Feature (MIBF) selection algorithm, and a Stacking Ensemble Classifier (SAEC) is adopted for unified view classification. To enhance clinical interpretability, SHapley Additive exPlanations (SHAP) is applied. The proposed framework outperforms single-view feature learning methods by 3% on average and state-of-the-art techniques by 2% in classification accuracy, sensitivity, specificity, and F1-score using the CHB-MIT Scalp and Bonn EEG datasets. This approach offers an effective tool for EEG-based seizure detection (ESD) in clinical and healthcare settings.

Objective.Despite decades of electroencephalography (EEG) research, the relationship between EEG and underlying spiking dynamics remains unclear. This limits our ability to infer neural dynamics reflected in intracranial signals from EEG, a critical step to bridge electrophysiological findings across species and to develop non-invasive brain-machine interfaces (BMIs). In this study, we aimed to estimate spiking activity in the visual cortex using non-invasive scalp EEG.Approach. We recorded spiking activity from a 32-channel floating microarray permanently implanted in parafoveal V1 and scalp-EEG in a male macaque monkey. While the animal fixated, the screen flickered at different temporal frequencies to induce steady-state visual evoked potentials. We analyzed the relationship between the V1 multi-unit spiking activity envelope (MUAe) and EEG frequency bands to predict MUAe at each time point from EEG. We extracted instantaneous spectrotemporal features of the EEG signal, including phase, amplitude, and phase-amplitude coupling of its frequency bands.Main results. Although the relationship between these spectrotemporal features and the V1 MUAe was complex and frequency-dependent, they were reliably predictive of the MUAe. Specifically, in a linear regression predicting MUAe from EEG, each EEG feature (phase, amplitude, coupling) contributed to model predictions. In addition, we found that MUAe predictions were better in shallow than deep cortical layers, and that the phase of stimulus frequency further improved MUAe predictions.Significance.Our study shows that a comprehensive account of spectrotemporal features of non-invasive EEG provides information on underlying spiking activity beyond what is available when only the amplitude or phase of the EEG signal is considered. This demonstrates the richness of the EEG signal and its complex relationship with neural spiking activity and suggests that using more comprehensive spectrotemporal signatures could improve BMI applications.

Long-term, continuous health monitoring imposes stringent demands on bio-recording analog front-end (AFE) circuits, specifically in terms of dynamic range (DR), noise, input impedance, and power consumption. This work introduces a DR-enhanced direct-digitization AFE based on a Δ-modulated trans-conductor (TC) stage, followed by a second-order ΔΣ ADC. In this architecture, the accumulated DAC is subtracted exclusively at the TC input stage, allowing the integrators to process only the low-amplitude Δ-modulated signal and thus relaxing the dynamic range constraints of conventional Gm-C ΔΣ ADCs. The TC input stage achieves high input impedance and high linearity through a current-balancing transconductor and a flipped-voltage-follower (FVF) loop. Fabricated with a standard 180nm CMOS process, the proposed Δ-ΔΣ AFE exhibits an SNDR of 91 dB, a dynamic range of 101 dB, input referred noise of 58 nV/$\surd{\rm Hz}$, and a power consumption of 63 $\boldsymbol{\mu}$W. These results correspond to a FoMSNDR of 160.1 dB and a FoMDR of 170 dB. The AFE prototype has been validated through scalp EEG, leg EMG, and chest ECG with significant body movements, demonstrating its effectiveness as a motion-artifact-tolerant direct-ADC front end.

Evaluating the clinical readiness of artificial intelligence in EEG-based epilepsy diagnosis

Bridging surface and depth: A systematic review of seizure patterns in simultaneous scalp and stereo-EEG.

BackgroundLate‐onset unexplained epilepsy (LoUE), defined as epilepsy starting after age 55 with no clearly identified cause, has emerged as a significant risk factor for dementia. Individuals presenting with LoUE have no prior history of dementia. Yet, LoUE is associated with a 2‐3x increased risk of developing dementia, and up to 25% of individuals with LoUE develop dementia within 4 years after their first seizure. We have little understanding of the mechanisms that underlie development of dementia in LoUE.MethodThe ELUCID Study (Epilepsy of Late‐onset Unknown etiology as a risk factor for Cognitive Impairment and Dementia) is a multi‐center, prospective longitudinal observational study of LoUE, focused on understanding mechanisms and predicting outcomes of mild cognitive impairment and dementia in LoUE. ELUCID will enroll 600 participants with LoUE (and without dementia) across 7 study sites. Participants undergo a baseline evaluation with clinical history, cognitive testing, brain MRI, overnight scalp EEG, and blood draw, and are followed longitudinally with interval history every 6 months and annual cognitive testing. The primary outcomes are development of mild cognitive impairment and dementia.ResultTo date, 67 ELUCID participants have completed their initial study visit, with mean age of 67.9±7.2 years and 38.8% female. The sample includes 89.6% White, 3% Black, 1.5% Asian, 6% unreported race, and 1.5% Hispanic ethnicity. Mean level of education was 16.9±2.7 years. Vascular risk factors were common, including hypertension (51%), hyperlipidemia (58%), diabetes mellitus (6%), coronary artery disease (9%), and obstructive sleep apnea (28%). A family history of seizures was present in 23.9% of participants, and a family history of dementia in 58%. Cognitive test scores largely fell within normal range, including: MMSE: 28.7±1.5; Logical Memory Delayed: 11.9±3.4; FCSRT Free Recall: 31.6±6.4; Trails B: 94.3±54.4; Digit Symbol Substitution: 41.9±10.1; and Category Fluency (animals): 17.0±4.9. Subjectively, 32.8% of participants felt their memory had worsened compared to 6 months prior.ConclusionThe ELUCID Study is a large longitudinal study of LoUE that will define its relationship to Alzheimer's disease and related dementias. Here we describe the study protocol and provide an early report of the baseline demographic and clinical characteristics of the accruing ELUCID study population.

Interictal epileptiform discharges (IEDs) are crucial for epilepsy diagnosis but are often undetectable on scalp EEG (scEEG). This study aimed to develop a Support Vector Machine classifier to detect mesial temporal lobe (MTL) IEDs invisible on scEEG using simultaneous intracranial EEG (iEEG) and scEEG recordings and to identify contributing scEEG features. Data from 17 patients with epilepsy were analyzed. IED epochs were extracted where IEDs were present on iEEG but absent on scEEG. Control epochs were selected from periods without IEDs on both EEGs. Feature selection was performed, and a classifier was developed and validated with external data from 35 MTL epilepsy (MTLE) and 33 non-epileptic patients. The classifier used 58 selected features and achieved an accuracy of 0.70 and an area under the receiver operating characteristic curve of 0.78 on holdout validation. External validation revealed significant differences in IED-classified epoch frequencies before and after drug withdrawal in patients with MTLE, and between MTLE and non-epileptic groups. Feature analysis identified high-frequency power suppression, increased ipsilateral connectivity, and enhanced cross-frequency coupling as markers of IEDs that were undetectable in scEEG . This study shows that machine learning can detect MTL IEDs invisible to scEEG, revealing related scEEG changes and aiding EEG analysis.