Journal of neural engineering最新文献

{"title":"Temporal basis function models for closed-loop neural stimulation.","authors":"Matthew J Bryan, Felix Schwock, Azadeh Yazdan-Shahmorad, Rajesh P N Rao","doi":"10.1088/1741-2552/ae036a","DOIUrl":"https://doi.org/10.1088/1741-2552/ae036a","url":null,"abstract":"<p><p>Closed-loop neural stimulation provides novel therapies for neurological diseases such as Parkinson's disease (PD), but it is not yet clear whether artificial intelligence (AI) techniques can tailor closed-loop stimulation to individual patients or identify new therapies. Further advancements are required to address a number of difficulties with translating AI to this domain, including sample efficiency, training time, and minimizing loop latency such that stimulation may be shaped in response to changing brain activity.</p><p><strong>Approach: </strong>we propose temporal basis function models (TBFMs) to address these difficulties, and explore this approach in the context of excitatory optogenetic stimulation. We demonstrate the ability of TBF models to provide a single-trial, spatiotemporal forward prediction of the effect of optogenetic stimulation on local field potentials (LFPs) measured in two non-human primates. The simplicity of TBF models allow them to be sample efficient (<20min of training data), rapid to train (<5min), and low latency (<0.2ms) on desktop CPUs.</p><p><strong>Main results: </strong>we demonstrate the model on 40 sessions of previously published excitatory optogenetic stimulation data. Surprisingly, on test sets it achieved a prediction accuracy 44% higher than a complex nonlinear dynamical systems model that requires hours to train, and 158% higher than a linear state-space model requiring 90 minutes to train. Additionally, in two simulations we show that it successfully allows a closed-loop stimulator to drive neural trajectories, and to achieve the user-preferred trade-offs between under- and over-stimulation, given the uncertainty in the model; it achieves an area under curve (AUC) of 0.7 in both cases.</p><p><strong>Significance: </strong>by optimizing for sample efficiency, training time, and latency, our approach begins to bridge the gap between complex AI-based approaches to modeling dynamical systems and the vision of using such forward prediction models to develop novel, clinically useful closed-loop stimulation protocols.</p>","PeriodicalId":94096,"journal":{"name":"Journal of neural engineering","volume":" ","pages":""},"PeriodicalIF":3.8,"publicationDate":"2025-09-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145002379","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Extracting robust single-trial somatosensory evoked potentials for non-invasive brain computer interfaces.","authors":"Disha Gupta, Jodi Brangaccio, Helia Mojtabavi, Jonathan Wolpaw, N Jeremy Hill","doi":"10.1088/1741-2552/adfd8a","DOIUrl":"10.1088/1741-2552/adfd8a","url":null,"abstract":"<p><p><i>Objective.</i>Reliable extraction of single-trial somatosensory evoked potentials (SEPs) is essential for developing brain-computer interface (BCI) applications to support rehabilitation after brain injury. For real-time feedback, these responses must be extracted prospectively on every trial, with minimal post-processing and artifact correction. However, noninvasive SEPs elicited by electrical stimulation at recommended parameter settings (0.1-0.2 msec pulse width, stimulation at or below motor threshold, 2-5 Hz frequency) are typically small and variable, often requiring averaging across multiple trials or extensive processing. Here, we describe and evaluate ways to optimize the stimulation setup to enhance the signal-to-noise ratio (SNR) of noninvasive single-trial SEPs, enabling more reliable extraction.<i>Approach.</i>SEPs were recorded with scalp electroencephalography in tibial nerve stimulation in thirteen healthy people, and two people with CNS injuries. Three stimulation frequencies (lower than recommended: 0.2 Hz, 1 Hz, 2 Hz) with a pulse width longer than recommended (1 msec), at a stimulation intensity based on H-reflex and M-wave at Soleus muscle were evaluated. Detectability of single-trial SEPs relative to background noise was tested offline and in a pseudo-online analysis, followed by a real-time demonstration.<i>Main</i><i>results.</i>SEP N70 was observed predominantly at the central scalp regions. Online decoding performance was significantly higher with Laplacian filter. Generalization performance showed an expected degradation, at all frequencies, with an average decrease of 5.9% (multivariate) and 6.5% (univariate), with an AUC score ranging from 0.78-0.90. The difference across stimulation frequencies was not significant. In individuals with injuries, AUC of 0.86 (incomplete spinal cord injury) and 0.81 (stroke) was feasible. Real-time demonstration showed SEP detection with AUC of 0.89.<i>Significance.</i>This study describes and evaluates a system for extracting single-trial SEPs in real-time, suitable for a BCI-based operant conditioning. It enhances SNR of individual SEPs by alternate electrical stimulation parameters, dry headset, and optimized signal processing.</p>","PeriodicalId":94096,"journal":{"name":"Journal of neural engineering","volume":"22 5","pages":""},"PeriodicalIF":3.8,"publicationDate":"2025-09-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12405709/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144984201","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Disrupted periodic spatiotemporal pattern and dynamic reorganization of its basic states in generalized epilepsy.","authors":"Junxia Chen, Sisi Jiang, Guofeng Ye, Zhihuan Yang, Changyue Hou, Hechun Li, Haonan Pei, Roberto Rodriguez-Labrada, Jianfu Li, Dezhong Yao, Cheng Luo","doi":"10.1088/1741-2552/ae01d9","DOIUrl":"https://doi.org/10.1088/1741-2552/ae01d9","url":null,"abstract":"<p><strong>Objective: </strong>The anti-correlation between the default mode network (DMN) and task-positive network (TPN) is a stable characteristic of normal brain activity. However, in idiopathic generalized epilepsy (IGE), this anti-correlation is often disrupted and strongly associated with epileptic seizures. This study aims to use periodic spatiotemporal patterns (PSTP) analysis to elucidate the relationship between the DMN-TPN anti-correlation and epileptic activity, providing new insights into the neural mechanisms underlying IGE.</p><p><strong>Approach: </strong>Resting-state functional magnetic resonance imaging was used to analyze PSTP in both healthy controls (HC) and IGE patients. A pattern-finding algorithm was initially applied to identify repeated spatiotemporal patterns, followed by a novel PSTP-finding algorithm to uncover dynamic periodic patterns through analysis of fluctuations in the DMN-TPN anti-correlation. The Hilbert transform was applied to capture the underlying basic states of these periodic patterns. Additionally, the relationship between period length and intrinsic neural timescales (INT) was explored.</p><p><strong>Main results: </strong>IGE patients exhibited a reduced DMN-TPN anti-correlation, particularly during TPN-dominant states. Additionally, IGE patients exhibited greater dynamic instability in basic states, marked by more frequent transitions between transitional states. Furthermore, lower correlations between period length and INT were observed in cognitive regions of IGE patients.</p><p><strong>Significance: </strong>These findings suggest that dynamic switching between the DMN and TPN in IGE is weaker and less balanced, with disruptions in periodic rhythms linked to cognitive impairments. The proposed PSTP framework provides new insights into the abnormal rhythms of IGE from a spatiotemporal perspective.</p>","PeriodicalId":94096,"journal":{"name":"Journal of neural engineering","volume":" ","pages":""},"PeriodicalIF":3.8,"publicationDate":"2025-09-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144983770","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Closed-form expressions for the directions of maximum modulation depth in temporal interference electrical brain stimulation.","authors":"Mariano Fernández Corazza, Sergei Turovets, Carlos Horacio Muravchik","doi":"10.1088/1741-2552/ae01dc","DOIUrl":"https://doi.org/10.1088/1741-2552/ae01dc","url":null,"abstract":"<p><strong>Objective: </strong>In temporal interference (TI) transcranial electrical stimulation (tES), an emerging brain stimulation technique, the interference of two high-frequency currents with a small frequency difference is used to target specific brain regions with better focality than in standard tES. While the magnitude of the modulation depth has been previously investigated, an explicit formula for the direction in which this modulation is maximized has been lacking. This work provides a novel closed-form analytical expression for the orientation of maximum modulation depth in TI tES. We also found a secondary orientation where the modulation depth has a local maximum. Moreover, we provide closed-form analytical formulas for this orientation as well as for the modulation depth along this orientation. To our knowledge, these closed-form expressions and the presence of the secondary maximum have not been previously reported.</p><p><strong>Approach: </strong>We derive compact analytical expressions and validate them through comprehensive computational simulations using a realistic human head model. We also provide a complete analytical derivation of the widely used formula for the maximum modulation depth magnitude stated in Grossman et al, 2017.</p><p><strong>Main results: </strong>Our simulations demonstrate that the modulation depth predicted with our new analytical direction formula is indeed the maximum compared to other directions. The derived closed-form expression provides a faster and more accurate alternative to iterative numerical optimization methods used in previous studies to estimate this direction. Furthermore, we found that due to interference in 3D, the modulation depth along the secondary maximum orientation can be of similar strength to the maximum modulation depth intensity when interfering electric field vectors are significantly misaligned. Finally, we show that by modifying the ratio of the injected current strengths, it is possible to steer these directions and fine-tune the stimulation along a desired direction of interest.</p><p><strong>Significance: </strong>Overall, this work provides a detailed treatment of TI electric fields in 3D. The presented closed-form expressions for the directions of maximum and secondary maximum modulation depths are relevant for the better interpretation of both simulated and experimental results in TI studies by allowing comparison with neuronal orientations in the brain.</p>","PeriodicalId":94096,"journal":{"name":"Journal of neural engineering","volume":" ","pages":""},"PeriodicalIF":3.8,"publicationDate":"2025-09-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144983780","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Instantaneous recognition method for lower limb continuous motion based on onset-window surface electromyography data.","authors":"Xiaohui Li, Hao Zhou, Xueyan Lyu, Xiaoyue Yu, Dezhi Yu, Wenzhuo Wang, Guanglin Li, Lin Wang","doi":"10.1088/1741-2552/adfab3","DOIUrl":"10.1088/1741-2552/adfab3","url":null,"abstract":"<p><p><i>Objective</i>. Human-robot collaboration in lower-limb rehabilitation devices imposes stringent requirements on both the recognition accuracy of motion intention and real-time responsiveness. The precise recognition of lower limb motion based on surface electromyography (sEMG) has always been a primary focus of study. However, achieving low-delay recognition in lower limb continuous motion while maintaining accuracy remains a challenge, which is key to unlocking the full potential for the effective deployment and widespread application of robots.<i>Approach</i>. An innovative recognition method in lower limb continuous motion was presented in this paper, which investigated the instantaneous recognition network (IRN) and continuous recognition (CR) model.<i>Main results.</i>The comparative analysis revealed that by utilizing an optimal length of 210 for the onset-window sEMG data, the proposed IRN could substantially reduce the time delay from 300 ms/350 ms to 60 ms at the methodological level. The implementation of the class-balanced method enhanced motion recognition accuracy by an additional 4.83% within the onset window. The CR model was validated across seven scenarios, comprehensively covering all potential situations in daily continuous movements, and achieved an average accuracy of 96.31%.<i>Significance.</i>This study demonstrates the potential of the proposed instantaneous recognition method to enhance performance in lower limb continuous motion, providing an innovative approach for research on human-robot synchronization.</p>","PeriodicalId":94096,"journal":{"name":"Journal of neural engineering","volume":" ","pages":""},"PeriodicalIF":3.8,"publicationDate":"2025-09-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144839437","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Motor unit number estimation based on convolutional neural network.","authors":"Chen Junjun, Zezhou Li, Linyan Wu, Zhiyuan Lu, Maoqi Chen, Ping Zhou","doi":"10.1088/1741-2552/ae01da","DOIUrl":"https://doi.org/10.1088/1741-2552/ae01da","url":null,"abstract":"<p><strong>Objective: </strong>The compound muscle action potential (CMAP) scan contains a muscle's detailed stimulus-activation information and thereby can be used for motor unit number estimation (MUNE). Due to the challenges in accurately obtaining the motor unit numbers from experimental CMAP scans, most existing MUNE methods rely on data fitting, which is time-consuming and requires manual operations. This study explored the feasibility of a neural network-based MUNE approach and proposes an end-to-end model for rapid estimation. &#xD;Approach. We developed NNEstimation, a novel supervised learning framework based on a convolutional neural network (CNN), to estimate motor unit numbers from both synthetic and experimental CMAP scans. A probabilistic model with varied parameters was used to generate CMAP scans with diverse characteristics for neural network training. NNEstimation trained on synthetic data was directly tested on both synthetic and experimental data. &#xD;Main Results. Evaluations on synthetic CMAP scans demonstrate that NNEstimation achieves lower estimation error and shorter execution time than the conventional data fitting method, MScanFit. The accuracy of NNEstimation is influenced by motor unit numbers of CMAP scans and remains unaffected by noise levels, recorded amplitudes, or motor unit activation thresholds. Moreover, NNEstimation's estimates on experimental data are highly consistent with those of MScanFit. &#xD;Significance. Although trained solely on synthetic CMAP scans, NNEstimation achieves estimation results comparable to those of the traditional algorithm on experimental data while significantly reducing execution time, demonstrating its potential for practical MUNE applications.</p>","PeriodicalId":94096,"journal":{"name":"Journal of neural engineering","volume":" ","pages":""},"PeriodicalIF":3.8,"publicationDate":"2025-09-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144984072","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Refining signal localization: a transcranial magnetic stimulation motor mapping approach to functional near-infrared spectroscopy.","authors":"Garrett S Black, Joshua A Dugdale, Jordan A Borrell","doi":"10.1088/1741-2552/adfd8b","DOIUrl":"https://doi.org/10.1088/1741-2552/adfd8b","url":null,"abstract":"<p><p><i>Objective.</i>Functional near-infrared spectroscopy (fNIRS) has emerged as a promising brain imaging tool due to its cost-effectiveness and balance between spatial and temporal resolution. However, its reliance on the 10-20 EEG coordinate system for probe placement introduces potential inaccuracies in cortical localization. Despite concerns regarding its spatial precision, the integration of transcranial magnetic stimulation (TMS) with fNIRS for validating signal localization has not been systematically explored. This study aimed to demonstrate the interindividual variability in hand motor representations and how it influences the precision of fNIRS recordings during motor tasks.<i>Approach.</i>Neuronavigated TMS was employed on 18 neurotypical adults to map the motor representations of the first dorsal interosseous (FDI) and fourth dorsal interosseous (4DI) muscles. Center-of-gravity (CoG) coordinates from TMS-evoked motor maps were compared with fNIRS channel locations, including the theoretical hand channel defined by the 10-20 EEG system. FNIRS signals were recorded during a hand-grasp motor task, and the subject-specific hand channel was determined by identifying the fNIRS channel closest to the individual's TMS CoG.<i>Main Results.</i>TMS motor mapping revealed substantial interindividual variability, with 56% of participants demonstrating deviations from the theoretical fNIRS hand channel. TMS motor maps showed that the FDI and 4DI representations were closely positioned, with the 4DI representation slightly anterior to the FDI (<i>p</i>= 0.022). Analysis of fNIRS signals indicated that subject-specific hand channels exhibited significantly higher hemodynamic response amplitudes compared to the theoretical hand channel (<i>p</i>= 0.0004), suggesting enhanced signal sensitivity when using individualized cortical mapping. Additionally, fNIRS signal variance was significantly higher in the theoretical channel, indicating greater signal variability and lower signal robustness.<i>Significance.</i>These findings highlight the limitations of rigidly applying the 10-20 EEG system for spatial localization in fNIRS-based motor studies and show the benefits of integrating TMS-derived cortical mapping for improved signal accuracy and robustness.</p>","PeriodicalId":94096,"journal":{"name":"Journal of neural engineering","volume":"22 5","pages":""},"PeriodicalIF":3.8,"publicationDate":"2025-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144984179","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Towards stimulation-free automatic electrocorticographic speech mapping in neurosurgery patients.","authors":"Alexei Voskoboynikov, Magomed Aliverdiev, Yulia Nekrasova, Ilia Semenkov, Anastasia Skalnaya, Mikhail Sinkin, Alexei Ossadtchi","doi":"10.1088/1741-2552/adfc9c","DOIUrl":"10.1088/1741-2552/adfc9c","url":null,"abstract":"<p><p><i>Objective.</i>The precise mapping of speech-related functions is crucial for successful neurosurgical interventions in epilepsy and brain tumor cases. Traditional methods like electrocortical stimulation mapping (ESM) are effective but carry a significant risk of inducing seizures.<i>Methods.</i>To address this, we have prepared a comprehensive ESM + electrocorticographic mapping (ECM) dataset from 14 patients with chronically implanted stereo-EEG electrodes. Then we explored several compact machine learning (ML) approaches to convert the ECM signals to the ground truth derived from the risky ESM procedure. Both procedures involved the standard picture naming task. As features, we used gamma-band power within successive temporal windows in the data averaged with respect to picture and voice onsets. We focused on a range of classifiers, including XGBoost, linear support vector classification (SVC), regularized logistic regression, random forest,<i>k</i>-nearest neighbors, decision tree, multi-Layer perceptron, AdaBoost and Gaussian Naive Bayes classifiers and equipped them with confidence interval estimates, crucial in a real-life application. We validated the ML approaches using a leave-one-patient-out procedure and computed ROC and Precision-Recall curves for various feature combinations.<i>Results.</i>For linear SVC we achieved ROC-AUC and PR-AUC scores of 0.91 and 0.88, respectively, which effectively distinguishes speech-related from non-related iEEG channels. We have also observed that the use of information on the voice onset moment notably improved the classification accuracy.<i>Significance.</i>We have for the first time rigorously compared the ECM and ESM results and mimicked a real-life use of the ECM technology. We have also provided public access to the comprehensive ECM+ESM dataset to pave the road towards safer and more reliable eloquent cortex mapping procedures.</p>","PeriodicalId":94096,"journal":{"name":"Journal of neural engineering","volume":" ","pages":""},"PeriodicalIF":3.8,"publicationDate":"2025-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144877647","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Deep brain stimulation-induced local evoked potentials outperform spectral features in spatial and clinical STN mapping.","authors":"Enrico Opri, Faical Isbaine, Seyyed Bahram Borgheai, Emily Bence, Roohollah Jafari Deligani, Jon T Willie, Robert E Gross, Nicholas Au Yong, Svjetlana Miocinovic","doi":"10.1088/1741-2552/adf99f","DOIUrl":"10.1088/1741-2552/adf99f","url":null,"abstract":"<p><p><i>Objective.</i>Deep brain stimulation (DBS) of the subthalamic nucleus (STN) is an established therapy for Parkinson's disease (PD). Yet, optimizing lead placement and stimulation programming remains challenging. Current techniques rely on imaging and intraoperative microelectrode recordings (MER), while programming relies on trial-and-error clinical testing, which can be time-consuming. DBS-induced local evoked potentials (DLEP), also known as evoked resonant neural activity, have emerged as a potential alternative electrophysiological marker for mapping. However, direct comparisons with traditional spectral features, such as beta-band, high-frequency oscillations (HFOs), and aperiodic component are lacking.<i>Approach.</i>We evaluated DLEP across 39 STN DBS leads in 31 subjects with PD undergoing DBS surgery, using both a single-pulse and high-frequency (HF) burst stimulation paradigms. We developed a novel artifact-removal method to enable monopolar DLEP recovery, including estimating the DLEP amplitudes at stimulated contacts, further enhancing spatial sampling of DLEP. We evaluated spectral features and DLEP in respect to imaging-based and MER-based localization, and its predictive power for post-operative programming.<i>Main results.</i>DLEP showed great spatial consistency, maximizing within STN with 100% accuracy for single-pulse and 84.62% for burst stimulation, surpassing spectral measures including beta (89.74%) and HFO (82.05%). DLEP better correlated with clinical outcomes (single-pulses<i>ρ</i>= -0.33, HF bursts<i>ρ</i>= -0.26), than spectral measures (beta<i>ρ</i>= -0.25, HFO<i>ρ</i>= 0.05). Furthermore, single-pulses at low-frequencies are sufficient for DLEP-based mapping.<i>Significance.</i>We show how DLEP provide higher STN-spatial specificity and correlation with postoperative programming compared to spectral features. To support clinical translation of DLEP, we developed two methods aimed to recover artifact-free DLEP and estimating DLEP amplitudes at stimulating contacts. DLEP appear distinct from beta and HFO activity, yet strongly tied to aperiodic spectral components, suggesting that DLEP amplitude reflects underlying STN excitability. This study highlights that DLEP are a robust and clinically valuable marker for DBS targeting and programming.</p>","PeriodicalId":94096,"journal":{"name":"Journal of neural engineering","volume":" ","pages":""},"PeriodicalIF":3.8,"publicationDate":"2025-08-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12395123/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144805526","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Feature fusion based on global-local weighted attention model for automatic epileptic seizure detection.","authors":"Xiang Li, Ke Zhang, Xin Wang, Zhiheng Zhang, Pengsheng Zhu, Mingxing Zhu, Xianhai Zeng, Shixiong Chen","doi":"10.1088/1741-2552/ae00f4","DOIUrl":"https://doi.org/10.1088/1741-2552/ae00f4","url":null,"abstract":"<p><strong>Objective: </strong>Epilepsy is a neurological disorder characterized by recurrent seizures, which present significant challenges in both diagnosis and treatment. Despite advances in seizure detection, existing methods often struggle with accurately capturing the complex and dynamic interactions between temporal, spatial, and spectral features of EEG signals. This leads to limitations in the detection accuracy and generalization across different datasets.</p><p><strong>Approach: </strong>To address these challenges, we propose GLWA (Global-Local Weighted Attention) model, which integrates temporal, spatial, and spectral features through a local-global attention mechanism. At the same time, GLWA effectively balances both global and local features, capturing comprehensive information from EEG signals to enhance seizure detection accuracy.</p><p><strong>Main results: </strong>Our proposed model achieves accuracy rates of 98.82% and 98.89% on the CHB-MIT and Siena datasets, respectively. These results demonstrate the model's capability to effectively integrate these features, resulting in improved detection performance.</p><p><strong>Significance: </strong>Furthermore, we visualize the model's decision-making process to gain insights into the attention distribution across different brain regions and spectraluency bands, further emphasizing GLWA's potential in seizure detection. This work demonstrates the model's superior performance and interpretability, providing a robust approach for accurate and generalizable identification of seizures.</p>","PeriodicalId":94096,"journal":{"name":"Journal of neural engineering","volume":" ","pages":""},"PeriodicalIF":3.8,"publicationDate":"2025-08-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144983723","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}