International IEEE/EMBS Conference on Neural Engineering : [proceedings]

Objective:Reliable detection of high-frequency oscillations (HFOs) in intracranial EEG (iEEG) is critical for localizing the seizure onset zone (SOZ). However, artifact contamination severely impairs detection specificity and contributes to false positives. Methods:We developed a reliability guided HFO analysis pipeline that targets artifact elimination through a multi-level strategy, including event level denoising, channel signal quality assessment, and segment level temporal reliability analysis. The classifier was trained on 22 subjects using manual annotations of reliable versus unreliable data and validated using leave-one-subject-out cross-validation (LOSO-CV). Results:The final pipeline was further applied to an additional cohort of 8 independent subjects to evaluate generalizability and clinical utility. SOZ localization accuracy based on HFOs retained after event-level denoising improved from 66% (baseline) to 82%, and further to 93% after applying channel and segment reliability filters. ROC analysis confirmed that reliability-based filtering enhanced separation between SOZ and non-SOZ channels, improving clinical specificity. Conclusion:Incorporating spatial and temporal contextual information, specifically by evaluating signal quality across neighboring channels and consistency across adjacent time windows, significantly enhances the ability to distinguish true HFOs from artifacts. This approach improves both the robustness and clinical accuracy of HFO-based SOZ localization.

Stroke often causes damage to the motor descending pathways in the lesioned hemisphere, resulting in compensatory recruitment of contralesional motor resources. This proof-of-concept study examines how motor task demand affects brain-muscle communication in individuals with chronic stroke, which represents motor system remodeling after the injury. Coherence between EEG and deltoid EMG was estimated during an isometric shoulder abduction task, and hemisphere dominance was evaluated using a Laterality Index (LI) in five stroke survivors and five healthy controls during isometric shoulder abduction at 20% and 40% of maximal voluntary contraction. Compared to controls, stroke participants exhibited a significantly reduced beta-band corticomuscular connectivity (CMC) at the lesioned hemisphere, with enhanced contralesional CMC. This was accompanied by an increase in motor demand, which amplified contralesional cortical involvement and revealed a pronounced shift in the Laterality Index (LI) toward contralesional dominance. This shift is indicative of maladaptive neural reorganization and diminished reliance on the ipsilesional hemisphere. The potential of beta-band CMC and LI as sensitive neurophysiological biomarkers for evaluating motor impairment and guiding individualized neurorehabilitation strategies that account for cortical plasticity and load-dependent recruitment patterns in stroke recovery is underscored by these task-dependent alterations in coherence and hemispheric dominance.

High-frequency oscillations (HFOs) are promising biomarkers for localizing the seizure onset zone (SOZ) in drug-resistant epilepsy. Automatic HFO detectors often depend on specific parameter choices, limiting effectiveness across diverse datasets. This study optimizes a published amplitude-threshold-based detector using a data-driven approach on intracranial EEG recordings from 20 patients at four epilepsy centers. We use the Tree-structured Parzen Estimator (TPE) to efficiently search the parameter space, guided by a new objective function, the SOZ Detection Quality (SDQ) score, which balances SOZ localization precision with the number of detected events in the SOZ. In the training cohort, the optimized detector increased the proportion of detections in SOZ channels from 49% to 61% and improved AUROC from 0.78 to 0.89 while largely preserving the SOZ detection counts. Similar improvements were observed in the test cohort, where detections in SOZ rose from 45% to 62%, and AUROC increased from 0.92 to 0.95. In both cohorts, detections outside the SOZ were significantly reduced, enhancing spatial specificity without compromising sensitivity. Theoretically, our TPE-based optimization achieved over 2,958×speedup over a grid search. These results show an effective data-driven parameter tuning approach that improves HFO biomarker reliability and supports their use in clinical planning for epilepsy surgery.

Deep learning has the potential for advancing EEG analysis and interpretation, however widespread implementation has been constrained by the need for extensive labeled datasets and limited transparency. In contrast, generative pretrained transformer (GPT) architectures have enabled efficient learning of small datasets through fine-tuning and also offer verifiable chain-of-thought reasoning in agentic frameworks. Here, we present EEG-GPT, a framework for EEG classification fine-tuned on publicly available large-language models. We benchmark EEG-GPT on the Temple University Hospital (TUH) EEG corpus in a few-shot regime. Utilizing only 2 % of the training data, EEG-GPT classifies normal versus abnormal EEG recordings with AUROC of 0.86, comparable to state-of-the-art deep learning methods. In addition, EEG-GPT demonstrates the capability to be used as an artificial intelligence (AI) agent, orchestrating usage of specialized EEG tools while providing step-by-step verifiability of its reasoning steps. These results underscore GPTs potential to advance EEG analysis and interpretation, highlighting their efficient learning in small data regimes and their potential to automate EEG analysis in a verifiable manner.

Transfer learning, a technique commonly used in generative artificial intelligence, allows neural network models to bring prior knowledge to bear when learning a new task. This study demonstrates that transfer learning significantly enhances the accuracy of sleep-stage decoding from peripheral wearable devices by leveraging neural network models pretrained on electroencephalographic (EEG) signals. Consumer wearable technologies typically rely on peripheral physiological signals such as pulse plethysmography (PPG) and respiratory data, which, while convenient, lack the fidelity of clinical electroencephalography (EEG) for detailed sleep-stage classification. We pretrained a transformer-based neural network on a large, publicly available EEG dataset and subsequently fine-tuned this model on noisier peripheral signals. Our transfer learning approach improved overall classification accuracy from 67.6% (baseline model trained solely on peripheral signals) to 76.6%. Notable accuracy improvements were observed across sleep stages, particularly lighter sleep stages such as REM and N1. These results highlight transfer learning's potential to substantially enhance the accuracy and utility of consumer wearable devices without altering existing hardware. Future integration of self-supervised learning methods may further boost performance, facilitating more precise, longitudinal sleep monitoring for personalized health applications.

Recent advancements in Large Language Models have inspired the development of foundation models across various domains. In this study, we evaluate the efficacy of Large EEG Models (LEMs) by fine-tuning LaBraM, a state-of-the-art foundation EEG model, on a real-world stress classification dataset collected in a graduate classroom. Unlike previous studies that primarily evaluate LEMs using data from controlled clinical settings, our work assesses their applicability to real-world environments. We train a binary classifier that distinguishes between normal and elevated stress states using resting-state EEG data recorded from 18 graduate students during a class session. The best-performing fine-tuned model achieves a balanced accuracy of 90.47% with a 5-second window, significantly outperforming traditional stress classifiers in both accuracy and inference efficiency. We further evaluate the robustness of the fine-tuned LEM under random data shuffling and reduced channel counts. These results demonstrate the capability of LEMs to effectively process real-world EEG data and highlight their potential to revolutionize brain-computer interface applications by shifting the focus from model-centric to data-centric design.

Brain–computer interfaces (BCIs) have the potential to enable individuals to interact with devices by detecting their intention from brain activity. A common approach to BCI is to decode movement intention from motor imagery (MI), the mental representation of an overt action. However, research-grade electroencephalogram (EEG) acquisition devices with a high number of sensors are typically necessary to achieve the spatial resolution required for reliable analysis. This entails high monetary and computational costs that make these approaches impractical for everyday use. This study investigates the trade-off between accuracy and complexity when decoding MI from fewer EEG sensors. Data were acquired from 15 healthy participants performing MI with a 64-channel research-grade EEG device. After performing a quality assessment by identifying visually evoked potentials, several decoding pipelines were trained on these data using different subsets of electrode locations. No significant differences (p = [0.18–0.91]) in the average decoding accuracy were found when using a reduced number of sensors. Therefore, decoding MI from a limited number of sensors is feasible. Hence, using commercial sensor devices for this purpose should be attainable, reducing both monetary and computational costs for BCI control.

Surgical neuromodulation through implantable devices allows for stimulation delivery to subcortical regions, crucial for symptom control in many debilitating neurological conditions. Novel closed-loop algorithms deliver therapy tailor-made to endogenous physiological activity, however rely on precise sensing of signals such as subcortical oscillations. The frequency of such intrinsic activity can vary depending on subcortical target nucleus, while factors such as regional anatomy may also contribute to variability in sensing signals. While artefact parameters have been explored in more 'standard' and commonly used targets (such as the basal ganglia, which are implanted in movement disorders), characterisation in novel candidate nuclei is still under investigation. One such important area is the brainstem, which contains nuclei crucial for arousal and autonomic regulation. The brainstem provides additional implantation targets for treatment indications in disorders of consciousness and sleep, yet poses distinct anatomical challenges compared to central subcortical targets. Here we investigate the region-specific artefacts encountered during activity and rest while streaming data from brainstem implants with a cranially-mounted device in two patients. Such artefacts result from this complex anatomical environment and its interactions with physiological parameters such as head movement and cardiac functions. The implications of the micromotion-induced artefacts, and potential mitigation, are then considered for future closed-loop stimulation methods.

Mental fatigue results in feelings of tiredness and decreases task performance and efficiency. It can be caused by cognitive underload (passive fatigue) or overload (active fatigue). Although fatigue detection models have been developed using electroencephalogram (EEG) recordings, it is unclear whether their predictions (derived from one task) are applicable to different tasks. We investigated if an EEG-based fatigue model previously trained on a passive fatigue-inducing driving task, can predict fatigue during an active fatigue-inducing task. We collected EEG while subjects performed a target-hitting task as a non-fatigued baseline, followed by mental arithmetic tasks of increasing complexity and mental load. Overall, the fatigue score from the passive fatigue model was significantly correlated with self-reported fatigue and sleepiness, which increased over the course of the experiment. The fatigue score was higher during simple mental arithmetic than the target-hitting task, demonstrating that the model can successfully differentiate non-fatigued and passive/low mental load situations. However, the fatigue score was higher during the simple (low mental load) than the complex (high mental load) mental arithmetic task, showing that the model cannot accurately detect active fatigue. Future fatigue models should be trained on a variety of both active and passive fatigue-inducing tasks to be more generalizable.

Virtual environments are often used in pre-wearing training and assessment of prosthetic control abilities. Here, we developed a virtual prosthetic hand training platform for evaluation of closed-loop control of grasp force. Biorealistic controllers emulated a pair of antagonistic muscles that actuated the thumb and index fingers of the hand. Surface electromyographic (sEMG) signals from a pair of antagonistic residual muscles drove the biorealistic controllers. Tactile forces from fingertip sensors were conveyed to amputees through evoked tactile sensations (ETS) elicited at the projected finger map (PFM) areas of the stump. A forearm amputee subject participated in force tracking or holding tasks using the virtual hand with residual muscle EMGs, or the contralateral intact hand. Root-mean-square error (RMSE) was used as outcome measure of motor performance. Results in this subject showed that the biorealistic controller enabled the virtual hand to track and maintain grasping forces. The best performance in both tasks was achieved by the contralateral intact hand with visual feedback. The roles of visual or tactile feedback in force tracking or maintaining were also assessed with the virtual hand. For force holding task, hybrid tactile and visual feedback with biorealistic control had a better performance than single visual or tactile feedback in terms of RMSE, success rate, and force variability. While in the force pursuing task, tactile feedback did not seem to add visual feedback in following the target force. The study suggests that training may be required for a novel virtual hand user to perceive and integrate multiple modalities of feedback information, so as to optimize the closed-loop control ability.

The classification of motor imagery in non-invasive brain-computer interface (BCI) is a challenge due to the high variation of brain evoked responses across users and the non-stationarity properties of the electroencephalography (EEG) signal. With different sessions from the same user, it is possible to find substantial differences that require the BCI system to be recalibrated. In clinical settings, it is therefore necessary to know when a system should be recalibrated or when the system should adapt itself to deal with the shifts in the signal, i.e., the covariate shift, and/or catch artefacts that deviate substantially from the original data distribution. In this paper, we propose to use density based one-class classifiers using distances based on the Riemannian geometry framework for assessing the distribution of the EEG signal in motor imagery BCI. We assess the performance of the algorithms with a database of 14 participants. The results show that sessions from the same person can be reliably detected using the proposed approach. We also assess how the one-class classifiers can be used to determine if it is necessary to run domain adaptation in the test phase. The results support the conclusion that the accuracy improves as the system is adapted to shifting domains in signals.

Language processing in the brain comprises complex neurophysiology involving multiple regions and multimodal functions like cognition, memory, speech, etc. Recent studies show language processing is not restricted within anatomically established cortical language regions, and there is considerable inter-person variability. Clinically intraoperative functional language mapping is critical for studying patient-specific language processing and identifying the eloquent cortex to retain language functionalities during resection surgeries. Cortical auditory evoked potential (CAEP) induced by different auditory stimuli can provide a profound understanding of language processing. In this study, we present a low-cost, low-latency microcontroller-based audio system to induce and analyze CAEP. The system is established using off the shelf components; a 32bit Teensy (3.2) microcontroller, Teensy (3) audio shield etc. which makes it easily replicable. The Teensy can generate digital trigger locked to auditory stimulus onset to interface the system with biosignal amplifiers to record and study neural data simultaneously. The presented hardware tools enable the analysis of neural signals aligned with the audio stimuli with sub-millisecond temporal precision. Implementation of the system can provide a flexible platform to observe and analyze real-time cortical CAEP in the clinical settings. The system also provides great customization opportunities to study not only CAEP but also complex language processing tasks.

Microelectrodes are desired to deliver more charges to neural tissues while under electrochemical safety limits. Applying anodic bias potential during neurostimulation is a known technique for charge enhancement. Here, we investigated the levels of charge enhancement with anodic bias potential in vitro and in vivo using a custom-designed portable neurostimulator. We immersed our custom microelectrode probe in saline and measured voltage transients in response to constant current stimulation with and without a 500 mV anodic bias potential. We then inserted the same microelectrode probe into the primary motor cortex of the rat brain and measured voltage transients with the same electronics. Results showed that the charge injection capacity of the activated iridium oxide microelectrode site (with 2000 µm 2 geometric surface areas (GSAs)) increased by the use of the anodic bias potentials in both in vitro and in vivo: from 10 nC/phase to 32 nC/phase for 200 µs pulse widths, and from 2 nC/phase to 8 nC/phase, respectively. Thus, the order of charge injection capacities of the four cases tested in this study is as follows (from the lowest to the highest): in vivo without anodic bias, in vivo with anodic bias, in vitro without anodic bias, and in vitro with anodic bias. This work also validated in vivo use of our new portable neurostimulator which received stimulation waveforms wirelessly.

Intracortical brain computer interfaces (iBCIs) decode neural activity from the cortex and enable motor and communication prostheses, such as cursor control, handwriting and speech, for people with paralysis. This paper introduces a new iBCI communication prosthesis using a 3D keyboard interface for typing using continuous, closed loop movement of multiple fingers. A participant-specific BCI keyboard prototype was developed for a BrainGate2 clinical trial participant (T5) using neural recordings from the hand-knob area of the left premotor cortex. We assessed the relative decoding accuracy of flexion/extension movements of individual single fingers (5 degrees of freedom (DOF)) vs. three groups of fingers (thumb, index-middle, and ring-small fingers, 3 DOF). Neural decoding using 3 independent DOF was more accurate (95%) than that using 5 DOF (76%). A virtual keyboard was then developed where each finger group moved along a flexion-extension arc to acquire targets that corresponded to English letters and symbols. The locations of these letter/symbols were optimized using natural language statistics, resulting in an approximately a $2\times$ reduction in distance traveled by fingers on average compared to a random keyboard layout. This keyboard was tested using a simple real-time closed loop decoder enabling T5 to type with 31 symbols at 90% accuracy and approximately 2.3 sec/symbol (excluding a 2 second hold time) on average.

The brain-computer interface based on steady-state visual evoked potential (SSVEP) has received increasing attention due to its high information transfer rate and low subject variation. A major challenge of current SSVEP-BCI is the uncomfortableness and fatigue induced by the strong visual flicker. Thus, it is of vital importance to optimize SSVEP stimuli for a better user experience. Reducing the pixel density of stimuli is a promising method to improve SSVEP. However, it remains unknown how the neural responses would be when faced with low-pixel density stimuli, and it is also unclear whether the corresponding decoding accuracy can be improved or not. Hereto, this study investigated neural responses induced by the stimuli with distinct pixel densities (1%, 10%, 20%, 60%, 100%) under both low (8Hz, 15Hz) and high frequencies (33Hz, 40Hz, 60Hz), responses from parietal-occipital area were recorded by functional near-infrared spectroscopy (fNIRS) and electroencephalo-gram (EEG) concurrently, aiming to have a better understanding of low-pixel-density-related responses. As a result, the behavioral performance showed that the comfort index inclined as the pixel density became lower. EEG and fNIRS signal analysis indicated that 20%-pixel induced larger EEG and fNIRS response than most densities in the low-frequency band. As to classification, comparing to the 100%, classification accuracy of 20%-pixel density classifies significantly better in low-frequency and high-frequency bands, whether in EEG, fNIRS, or hybrid. The maximum classification accuracy of 20%-density can reach 97.66% in hybrid binary classification, with 3.77% more than 100% density. This research provides a theoretical and technical basis for developing user-friendly SSVEP-BCI.

The role of high-frequency oscillations (HFO) has been established in a multitude of the brain functions such as retrieval and consolidation of memory. Moreover, HFOs have been identified as a biomarker for pathological brain conditions, including epileptogenicity. Therefore, there has been a continuous effort to reliably detect and characterize HFOs. Here, we present an unsupervised HFO detector using characteristics of signals in the time-frequency domain obtained by continuous wavelet transform. By using L1 normalization for continuous wavelet transform, we improved the detection of HFOs without the need to normalize time-frequency maps. The elimination of normalizing the time-frequency maps reduces the computational cost of the analysis. We used two different benchmark datasets available in the literature to validate our proposed automatic HFO detector. The results demonstrate that our detector outperforms other commonly available HFO detectors including those that use timefrequency maps. Our HFO detector shows superior performance especially when signal-to-noise ratio (SNR) is low. Moreover, our detector can simultaneously detect artifacts, physiological spikes, and provide useful information about the HFOs such as their dominant frequency of oscillation, their average amplitude and their duration. This information can later be utilized to stratify HFOs for further analysis. Changes in HFO characteristics may be utilized as biomarkers in pathological conditions such as posttraumatic epilepsy.

Chronic functionality of neural interfaces (NI) is hampered by the physiological response to foreign objects, in part due to the mismatch of mechanical properties between soft neural tissue and the rigid materials used in interface construction. Polymer-based NIs have emerged as a key new technology in the pursuit of chronically stable neural recording and stimulation, but most polymer NIs are bespoke devices developed as part of specific research missions; many researchers do not have access to polymer-based NIs technology and among those who do there is a severe lack of standardization in material, construction, packaging, and testing, leading to a lack of repeatability among datasets. Here we present the Polymer Implantable Electrode (PIE) Foundry, a shared-resource for fabricating and disseminating standardized polymer-based microelectrode arrays for use in NIs. The model is based on the successful shared prototyping concept developed for the field of semiconductor research. Professional staff, supported by the BRAIN Initiative funding and operating in cleanroom space provided by the University of Southern California, offer design, fabrication, packaging, and testing of polymer-based microelectrode arrays as a free service to academic and non-profit research groups. The core enabling technology is a standardized set of micromachining protocols applied to the biocompatible, thin-film polymer Parylene C. By leveraging this method, we produce microelectrode arrays of varied size, shape, channel count, and application, disseminating hundreds of arrays to 18+ research groups in our first three years of operation. By standardizing materials, fabrication, and packaging, we create repeatable and comparable devices and have built a library of shareable designs. Channel counts range from 2 to 64, electrode sizes range from 15 μm diameter to 1 mm, designs include penetrating neural probes, spinal paddle electrodes, surface arrays for electroencephalography, and peripheral nerve cuffs for recording and stimulation, animal models include songbird, mouse, rat, cat, and sheep. Here we present details of our organizational structure, fabrication and packaging methods, representative examples of ex vivo and in vivo electrode performance, and key results from the first three years of Foundry operation.

Precise timing prediction is the ability to estimate time in millisecond timescale, it can speed up behavior, optimize perception, benefit adaptive behaviors. Behavioral studies have demonstrated training can improve the performance of precise timing prediction. However, neural evidence is still lacking in describing how training changes the neural characteristics of precise timing prediction. This study designed a cue-tapping (task1)/ timing-in-mind (task2) experiment, collected behavioral and electroencephalogram (EEG) data of 24 subjects in both before and after training period. Sample entropy (SampEn) and Lempel-Ziv complexity (LZC) were calculated to measure EEG complexity, functional connectivity based on phase locking value was also involved. Consequently, behavioral results showed that error time declined and accuracy rate increased as training progressed. SampEn was much smaller after training in almost all frequency-bands in task1, and reduced in task2 as well. LZC showed a decreased tendency after training, but no statistical significance was found. Moreover, after training, much stronger functional connectivity was found in low frequency-band in both tasks. The results can shed light on the modeling of training and precise predictive timing.

Brain stimulation has become an important treatment option for a variety of neurological and psychiatric diseases. A key challenge in improving brain stimulation is selecting the optimal set of stimulation parameters for each patient, as parameter spaces are too large for brute-force search and their induced effects can exhibit complex subject-specific behavior. To achieve greatest effectiveness, stimulation parameters may additionally need to be adjusted based on an underlying neural state, which may be unknown, unmeasurable, or challenging to quantify a priori. In this study, we first develop a simulation of a state-dependent brain stimulation experiment using rodent optogenetic stimulation data. We then use this simulation to demonstrate and evaluate two implementations of an adaptive Bayesian optimization algorithm that can model a dynamically changing response to stimulation parameters without requiring knowledge of the underlying neural state. We show that, while standard Bayesian optimization converges and overfits to a single optimal set of stimulation parameters, adaptive Bayesian optimization can continue to update and explore as the neural state is changing and can provide more accurate optimal parameter estimation when the optimal stimulation parameters shift. These results suggest that learning algorithms such as adaptive Bayesian optimization can successfully find optimal state-dependent stimulation parameters, even when brain sensing and decoding technologies are insufficient to track the relevant neural state.

This paper, for the first time, compares the behaviors of nonlinear versus linear muscle networks in decoding hidden peripheral synergistic neural patterns during dynamic functional tasks. In this paper, we report a case study during which one healthy subject conducts a series of four lower limb repetitive tasks. Specifically, the paper focuses on tasks that involve the right knee joint, including walking, sit-to-stand, stepping, and drop-jump. Twelve muscles were recorded using the Delsys Trigno system. The linear muscle network was generated using coherence analysis, and the nonlinear network was generated using Spearman's correlation. The results show that the degree, clustering coefficient, and global efficiency of the muscle network have the highest value among tasks in the linear domain for the walking task, while a low linear synergistic network behavior for the sit-to-stand is observed. On the other hand, the results show that the nonlinear functional muscle network decodes high connectivity (degree) and clustering coefficient and efficiency for the sit-to-stand when compared with other tasks. We have also developed a two-dimensional functional connectivity plane composed of linear and nonlinear features and shown that it can span the lower-limb dynamic task space. The results of this paper for the first time highlight the importance of observing both linear and nonlinear connectivity patterns, especially for complex dynamic tasks. It should also be noted that through a simultaneous EEG recording (using Brain Vision System), we have shown that, indeed, cortical activity may indirectly explain highly-connected nonlinear muscle network for the sit-to-stand task, highlighting the importance of nonlinear muscle network as a neurophysiological window of observation beyond the periphery.