Objective: Transferring large-scale medical foundation models to specific clinical tasks remains challenging, particularly in multi-center scenarios with heterogeneous data distributions and privacy constraints. Existing adaptation strategies provide limited solutions for collaboratively adapting foundation models across institutions while preserving their transferable representations. Methods: We propose FedSAM-3D, a foundation model adaptation framework for multi-center medical image segmentation. Built upon the SAM-Med3D backbone, FedSAM-3D defines the collaborative optimization space within adapter parameters while keeping the pretrained backbone unchanged. Through federated optimization within this constrained adaptation space, our framework enables efficient cross-center knowledge aggregation without exchanging full model parameters, while allowing each client to adapt the foundation model to local medical data distributions. Results: FedSAM-3D was evaluated on multi-center abdominal organ and brain tumor segmentation datasets under federated adaptation and zero-shot evaluation settings. Across both tasks and multiple clinical datasets, FedSAM-3D generally outperformed ablation variants and existing segmentation methods, demonstrating improved adaptation performance and robustness across heterogeneous medical data distributions. Moreover, FedSAM-3D achieved improved generalization on unseen external datasets, including cross-modality evaluation, highlighting its ability to enhance the transferability of medical foundation models. Conclusion: FedSAM-3D provides an effective paradigm for federated transfer of medical foundation models, achieving improved adaptation performance and generalization while avoiding direct sharing of raw medical data across institutions. Significance: FedSAM-3D provides a parameter-efficient approach for transferring medical foundation models across institutions without directly sharing raw data, facilitating their potential deployment in diverse clinical environments. Our code is available at https://github.com/huavhuahua/FedSAM-3D.
Objective: The creation and implementation of control strategies that govern the motions of prosthetic hands are essential for enhancing operational efficiency and user satisfaction. This paper introduces an H-infinity ($\mathcal {H}_\infty$) controller employing Timoshenko beam theory, specifically designed for a tendon-driven soft continuum wrist associated with a prosthetic hand known as 'PRISMA HAND II'. This combined approach focuses on achieving robust control that can effectively manage uncertainties and disturbances, ensuring that the prosthetic wrist responds accurately to the user's intentions. Methods: By employing the Timoshenko modeling approach, kinematic and dynamic models of the soft wrist are established, which are used in $\mathcal {H}_\infty$ controller to compute required tendon forces for achieving desired hand movements. These tendon forces are used for computing output deflections of the wrist. Results: The proposed controller is compared with other controllers to analyse performance of the proposed controller. The proposed $\mathcal {H}_\infty$ controller performed better compared to a Neural Network (NN) based adaptive controller in terms of Root Mean Square Error (RMSE) and steady state error values. Experimental evaluation demonstrated the controller's effectiveness in regulating wrist motions during real-time. Conclusion: The implementation of the control strategy is crucial for advancing the capabilities of soft continuum prosthetics, allowing for more natural and intuitive movements that align with the needs of the user with lower computational effort and better accuracy of motions. Significance: Our study contributes to the ongoing progress in adaptable prosthetic technologies, establishing a basis for extensive use in cost-effective healthcare solutions employing an efficient controller.
OBJECTIVE:Deep learning (DL) has become state-of-the-art for blood glucose (BG) forecasting in type 1 diabetes (T1D). However, its black-box nature raises safety and reliability concerns regarding its use for therapeutic decision-support. This study aims to: (1) highlight potential risks associated with standard DL-based BG forecasting, and (2) address them with PhyNet, a physiology-constrained monotonic neural network. METHODS:Two large-scale datasets (T1DEXI and MetaboNet, 848 subjects in total) were used to develop PhyNet-which enforces physiological consistency through a multi-branch structure and weight constraints-and compare it against six DL baselines (convolutional, recurrent, and transformer-based architectures). Models predicted BG levels up to 90-minute ahead using continuous glucose monitoring (CGM) data, carbohydrate intake, and insulin dosing, and were assessed for: (i) predictive accuracy with standard metrics, and (ii) adherence to physiological principles (i.e., carbohydrates increase BG, insulin lowers it) using explainable AI. Specifically, we evaluated model-predicted responses to varying carbohydrate and insulin intakes, and generated counterfactual explanations to identify model-recommended actions for avoiding adverse events. RESULTS:At a 30-minute horizon, predictive accuracy was similar across models (RMSE: 19.39-21.00 mg/dL; Time Gain: 10.45-13.35 min). Despite this, only PhyNet consistently captured the physiological effects of carbohydrates and insulin, yielding 0% unsafe recommendations versus up to 64.3% for baselines. CONCLUSION:Standard DL models can achieve state-of-the-art performance while failing to respect physiology, posing clinical risk. PhyNet preserves accuracy while enhancing physiological fidelity, supporting safer integration into T1D technologies. SIGNIFICANCE:Evaluating physiological consistency alongside predictive accuracy is essential for responsible clinical translation of DL-based BG forecasting.
OBJECTIVE:We evaluate the practical identifiability and clinical utility of local spectral graph model (SGM) parameters estimated from resting-state magnetoencephalography (MEG) in drug-resistant epilepsy. METHODS:A coupled excitatory-inhibitory SGM was fitted to MEG power spectra across 159 brain regions in 20 patients with temporal lobe epilepsy who achieved seizure freedom following surgery. Identifiability was assessed via boundary saturation analysis and inter-parameter correlations across four frequency bands. Identifiable parameters were tested for seizure onset zone (SOZ) discrimination. RESULTS:Model fit was excellent (mean $r = 0.976$) and significantly exceeded a $1/f^\beta$ baseline ($p < 10^{-10}$). Gain parameters ($g_{ei}$, $g_{ii}$) were robustly estimable, whereas the excitatory time constant ($\tau _{e}$) showed 74% boundary saturation in broadband fits, reduced to ${\sim }3\%$ when restricted to 1-50 Hz. SOZ regions exhibited elevated $g_{ei}$ (Cohen's $d = +0.67$, FDR-corrected $p = 0.024$) and reduced $g_{ii}$ ($d = -0.55$, $p = 0.003$), with a composite biomarker achieving 2.6-fold improvement over chance. CONCLUSION:Gain parameters are robustly identifiable from clinical MEG and capture excitatory-inhibitory imbalance in the SOZ, whereas time constants require band-limited fitting. These findings motivate an identifiability-aware framework in which only parameters demonstrably constrained by data are interpreted. SIGNIFICANCE:This is the first systematic assessment of neural mass model parameter identifiability in clinical epilepsy MEG, establishing practical guidelines for biophysical parameter interpretation.
Objective: Endovascular EEG (eEEG) has emerged as a brain monitoring technique that offers a balance between signal fidelity and invasiveness. Endovascular electrodes match subdural recordings in bandwidth and signal-to-noise ratio in animal studies, however, their signal properties remain sparsely quantified in humans. This study evaluated eEEG signals from five human participants undergoing intracarotid amobarbital injection (Wada test), while simultaneous scalp and endovascular EEG were recorded. Methods: All signals were preprocessed with artifact rejection and independent component analysis (ICA). Power spectral density (PSD), imaginary coherence (ImagC), phase-locking value (PLV), and amplitude envelope correlation (AmpC) were computed to quantify signal quality and functional connectivity. Results: The eEEG signals exhibited approximately ×3.7 higher power than concurrent scalp EEG, and nearest endovascular-scalp electrode pairs showed consistently higher coupling across all participants (mean difference 4.9 percentage points, range 1.8-7.6% across individuals), with effects most pronounced at distances $< $30 mm. Conclusion: These findings support the feasibility of eEEG for neuromonitoring and demonstrate its potential for simple brain-computer interface (BCI) applications. Significance: This work provides quantitative measures of the signal power and correlation with scalp EEG, obtained directly in humans for a microcatheter-deliverable wire electrode, establishing human operating bounds for endovascular EEG as a minimally invasive neural interface.
Objective: To develop a deep learning model for predicting histotripsy focal shifts in the liver caused by acoustic aberrations for real-time treatment optimization. Methods: A modified VGG19 CNN regression model was trained using 12,870 scenarios derived from 243 segmented human CT volumes. Input to the model consisted of 6-channel maps representing the distance through specific tissue types (bone, air, fat, muscle, liver, total tissue) along transducer-to-focus rays. Ground truth focus locations were determined via acoustic simulations (k-Wave) based on the minimum pressure location. The network was trained to output predicted focal shifts relative to the geometric focus location. Accuracy was evaluated as the mean absolute deviation between CNN predictions and ground truth simulations. An ablation analysis determined dominant features. Results: The simulation predicted aberration-induced focal shifts ranging from -12.9 to 3 mm. Across five-fold cross-validation, the model predicted shifts with a mean absolute deviation (standard deviation) across the folds of 0.3 (0.3), 0.3 (0.3), 0.5 (0.5) mm in X, Y, and Z, respectively. The corresponding mean bias was 0.10 mm, 0.08 mm, and 0.04 mm. A single one-sided t-test confirmed the absolute prediction errors were significantly less than the 1 mm clinical margin (p $\mathbf {< }$ 0.001). Ablation analysis revealed fat, liver, and total tissue as dominant predictors. Notably, CNN inference time was 13.8 ms, compared to a median 10.3 (IQR 3.6) hours for acoustic simulations performed on a high-performance cluster. Conclusion: The proposed CNN accurately predicts aberration shifts in focus location comparable to acoustic simulations with millisecond-scale inference speeds, enabling real-time aberration correction.
Imagined speech decoding remains challenging in brain-computer interfaces (BCIs) due to the low signal-to-noise ratio and complex spatio-temporal-spectral structure of Electroencephalogram (EEG) data. Existing studies mainly rely on single-view features or simple fusion strategies, limiting their ability to capture diverse neural characteristics during speech imagery. To address this limitation, we propose a Multi-view Feature Contrastive Decoupling and Enhancement (MFCDE) framework that integrates multi-view feature construction, feature decoupling, and adaptive masking. Four complementary views, including temporal, frequency-domain, phase-locking value (PLV), and graph-theoretic features, are extracted to characterize speech imagery-related neural dynamics. The decoupling mechanism reduces cross-view redundancy while preserving the discriminative information of each view. Experiments show that MFCDE consistently outperforms existing baselines in classification performance and stability. The learned view-shared and view-specific representations further provide neurophysiological insights by revealing the complementary contributions of temporal, spectral, and connectivity-based EEG patterns to imagined speech discrimination, indicating that reliable decoding depends on the joint utilization of neural dynamics, oscillatory activity, and inter-regional functional interactions.
OBJECTIVE:Gas embolism (GE) is a potentially life-threatening condition with difficult in vivo investigation. The physical nature of the initial stages of GE makes in vitro abiotic microfluidics valuable for phenomenological studies. A versatile microfluidic system was used to mimic the microscale iatrogenic GE during gastroscopy, laparoscopy, and hyperbaric therapy under three scenarios: gas transfer through perforated or non-perforated microvasculature, and via tissue supersaturation. METHODS:Polydimethylsiloxane devices mimicking microvascular geometries, with, or without injuries, hosted the flow of artificial blood. Localized gas pressures comparable to those used in gastroscopy, laparoscopy, and hyperbaric therapy were applied on microvasculature using air, carbon dioxide or /nitrogen to mimic microscale GE. RESULTS:Air and carbon dioxide produced markedly different GE patterns. Air generated numerous small bubbles across the entire pressure range, and a distinct population of larger emboli at lower pressures. Conversely, carbon dioxide embolism occurred beyond a threshold pressure (23-58 mm Hg), but once initiated, complete blockage or continuous gas injection lasting tens of seconds occurred. CONCLUSION:Air embolism may occur in injured microvasculature at pressures within the lower range encountered during gastroscopy. Carbon dioxide embolism requires higher pressures, but once triggered, it can result in high-volume gas entry during laparoscopy. High pressure gas-supersaturated tissues can act as gas 'reservoirs' prolonging GE therapy. SIGNIFICANCE:These findings underscore the risk of air GE in gastroscopy and the need for pressure and flowrate control; possible carbon dioxide GE during laparoscopy warrants further assessment; and the duration of gas clearance during GE therapy needs to be considered.
Objective: The detection of seizures from the vagus nerve electroneurogram (VENG) is an emerging application allowing minimally invasive closed-loop vagus nerve stimulation for the treatment of epilepsy. Several VENG-specific front-end designs were presented earlier, but their specifications were approximate or arbitrary, leading to inefficient or sub-optimal designs in an applicative context. In this work, we analyze how front-end design choices and non-idealities impact seizure detection performance. The sensitivity analysis leads to the presentation of potentially optimal front-end designs with low power consumption. Methods: This work uses experimental VENG data from 8 rats, a behavioral model of the front-end circuits, and a seizure detection algorithm based on template matching. The studied front-end limitations are intrinsic noise, common-mode rejection, dynamic range, quantization resolution, and oversampling rate, studied individually in a sensitivity analysis. Optimal designs are then proposed based on sensitivity results. The performance of the seizure detection algorithm is characterized by metrics that are least affected by the small dataset size. Results: Noise and dynamic range are the main factors impacting algorithm performance. An optimal front-end design with 3-μVRMS noise, 6-bit quantization, and 2.1-μW power consumption is presented and achieves perfect seizure classification on the dataset. Conclusion: The presented analysis enables the optimization of front-end specifications and a significant reduction in power consumption. Significance: This work bridges the gap between the performance of a seizure detection algorithm and the circuit-level specifications for ultra-low-power integrated bio-interface design.
Adhesive capsulitis (AC) is a musculoskeletal disorder that causes significant shoulder pain and stiffness. However, timely assessment of severity is often hindered by labor-intensive clinical procedures and relies largely on subjective judgment. In this study, we propose an explainable AI system for automated classification of AC severity using a single Azure Kinect 3D depth camera. Marker less 3D skeletal data were collected from 221 participants while they performed key shoulder movements, including abduction, flexion, and internal/external rotation. The proposed framework integrates a Temporal Convolutional Network (TCN) to model local temporal structure and cycle-to-cycle variation, followed by Transformer encoders to capture global contextual relationships across movement cycles. In addition, attention-based Long Short-Term Memory (LSTM), Bidirectional Long Short-Term Memory (BiLSTM), and Gated Recurrent Unit (GRU) networks were implemented as deep sequential baselines. For abduction, flexion, and external rotation, all architectures achieved test accuracies of 0.88-0.92 and macro-F1 scores ≥ 0.84. Notably, for internal rotation, the best-performing model (Attention-GRU) achieved an accuracy of 0.81 and macro-F1 of 0.77. Attention weights and SHAP analysis provided complementary interpretability, highlighting discriminative kinematic patterns between healthy and AC groups. The proposed system demonstrates the feasibility of automated AC severity stratification and supports clinician-facing reporting for rehabilitation-oriented assessment.
OBJECTIVE:Non-invasive Brain-Computer Interfaces (BCIs) based on Code-Modulated Visual Evoked Potentials (c-VEPs) using electroencephalography (EEG) signals require robust classification algorithms. It is unclear whether the best approach is to use a similarity measure or to follow a discriminant method. METHODS:We propose a multiple-classifier binary convolutional Siamese (MCBCS) network for single-trial c-VEP decoding, in which the multi-class recognition problem is decomposed into a set of class-specific binary similarity-learning tasks. The proposed MCBCS framework is systematically compared against a single multi-class Siamese network, convolutional neural networks for 63-bit m-sequence reconstruction and direct classification, and conventional correlation-based and canonical correlation analysis approaches. The study also investigates distance-based decoding strategies and the effect of temporal data augmentation with small to medium time shifts. RESULTS:Experimental results on EEG data from 13 subjects demonstrate that the MCBCS architecture consistently outperforms other tested methods under within-subject evaluation, with a mean single-trial accuracy of 96.89%. However, the MCBCS approach achieves 96.17% under a leave-one-subject-out protocol, while EEGNet achieves 96.79%. Finally, the Wasserstein Distance (WD$_{1}$) achieved the highest accuracy (93.88%) among the distance metrics. CONCLUSION:The multiple-classifier convolutional binary Siamese network achieved the highest overall performance. SIGNIFICANCE:The results highlight the effectiveness of class-specific similarity learning for robust compared to direct discriminant approaches.
Objective: This study presents an active sensing framework for information-optimal, model-based tracking of de formable anatomy to track the three-dimensional (3D) surface of the heart ventricles using a time sequence of two-dimensional (2D) image slices. Methods: A low-order deformable model parameterizes the cardiac surface, while cardiac motion dynamics are modeled by a recursive adaptive filter and tracked using a particle filter. Measurements of the system state are obtained from a magnetic resonance imaging (MRI) system whose slice selection is governed by an entropy-minimizing active sensing method that chooses the sensing action maximizing expected information gain. Performance is compared with cases where the image slice is fixed or randomly selected. The framework operates on general 2D image streams and is validated on multi-slice cine MRI data to enable comparison against ground-truth slice locations. A variable temporal sampling strategy reduces computational load by executing tracking updates at intervals defined by a temporal sampling factor. Results: The active sensing-based tracking method captured left ventricular (LV) surface points-of-interest within 3 pixels of accuracy at a mean root-mean-square error (RMSE) of 2.93mm, right ventricular (RV) tracking was 4.27mm, for an overall mean RMSE of 3.25mm. For a downsampling factor of 4, the framework maintains an overall RMSE of 3.57mm, a modest degradation relative to fully sampled tracking. Conclusion: The proposed frame work enables a flexible trade-off between computational efficiency and tracking performance. Significance: This work demonstrates that information-driven measurement selection can enhance de formable anatomical tracking under sensing constraints.
Objective: The shoulder joint is pivotal for upper-limb kinematics, yet individuals with post-stroke hemiparesis often require external shoulder support because of impaired motor control, excessive gravitational loading, and the risk of glenohumeral complications. Existing rehabilitation exoskeletons still face challenges in combining compliant movement assistance with sufficient antigravity load support. Methods: This study proposes a Pneumatic Shoulder Exosuit (PSE) designed to provide compliant multi-DOF shoulder assistance and augment load-bearing capacity during rehabilitation. The dual functionality is enabled by two coupled pouch-based pneumatic actuator modalities configured into a bionic agonist/antagonist architecture. Device performance was evaluated through experiments with 10 healthy volunteers using kinematic assessment, standardized trajectory-tracking, and static overhead weight-holding tasks. In addition, 7 participants with post-stroke hemiparesis were enrolled to evaluate the feasibility and usability of the PSE during assisted shoulder tasks. Results: The experimental results showed that the PSE can provide gravity support with up to 24 Nm of torque. The device maintained the movement trajectory with a 96.5% similarity compared to unassisted. It also reduced agonist muscle activities (anterior, middle, and posterior deltoids and biceps brachii) by up to 65% with a 70% reduction in compensatory muscle (latissimus dorsi) activity, and a 45% reduction in peak lateral trunk tilt. The participants with post-stroke hemiparesis also provided positive feedback. Conclusion: These findings demonstrate that the PSE can combine load-bearing augmentation with kinematic transparency during shoulder assistance. Significance: The PSE suggests potential for compliant rehabilitation-oriented shoulder support in future clinical and home-based rehabilitation studies.
Objective, scalable assessment of executive function in Attention-Deficit/Hyperactivity Disorder (ADHD) is limited by subjective rating scales and single modality tests. We present a timing-characterized virtual-reality (VR) implementation of the Wisconsin Card Sorting Test (WCST) synchronized with EEG and oculomotor/cephalic sensing for objective ADHD assessment in children. Stimulus-onset timing was empirically validated with a dedicated photodiode measurement across screen locations and distractor conditions. One hundred and one children (7-15 years; 64 ADHD, 37 control) completed a tutorial and a WCST session with distraction and no-distraction blocks. Task performance, head rotation, eye movement, and EEG features including N2/P3 event related potentials (ERP) were extracted, and group and condition contrasts on the ERP time-courses were evaluated using non-parametric cluster-based permutation testing. Classification used eight machine learning algorithms and a deep neural network (DNN) under 5-fold cross-validation; the best early-fusion DNN (task performance + head rotation + eye movement + N2/P3) reached 83% accuracy, and the best ERP-only model reached 79% under distraction. Overall, a timing characterized VR-EEG oculomotor fusion system captures ADHD-related executive function differences through complementary multimodal integration, providing a scalable, objective adjunct to clinical ADHD screening with empirically quantified stimulus-timing performance.
Brain-computer interfaces (BCIs) hold significant promise in medical applications, particularly for detecting consciousness in patients with disorders of consciousness (DoC). However, conventional BCIs often rely on visual stimuli, which limits their accessibility for visually impaired patients. To expand the applicability of BCIs to a wider range of patients, this study introduces an advanced auditory BCI system. The system incorporates an auditory paradigm utilizing stimulus-related semantic backgrounds and an improved EEG-Inception prototype network with ECA (ECAEI-ProNet). To validate the effectiveness of the proposed system, Experiment 1 was conducted to compare it against three control conditions: Condition 1, which included related backgrounds and stimuli; Condition 2, which included unrelated backgrounds and stimuli; and Condition 3, which presented no background. The results demonstrated that utilizing stimulus-related semantic backgrounds significantly improved BCI classification performance while eliciting event-related potentials (ERP) patterns associated with semantic consistency in subjects. Additionally, the two improvements made to the ECAEI-ProNet model, building upon EEG-Inception, enhanced the model's classification performance. Our proposed paradigm and classification model achieved online accuracies of 88.8$\pm$ 9.1%. To evaluate the clinical feasibility of the proposed BCI system, we applied it to 17 patients with DoC in Experiment 2. Results indicated that seven patients achieved online accuracy significantly above the chance level ($>$64%). Among them, the three highest-performing patients (P3, P5, and P12) showed improvements in both CRS-R scores and clinical diagnosis at the three-month follow-up. These findings suggest that the proposed auditory BCI system may offer preliminary information relevant to residual consciousness-related processing in some patients with DoC.
Objective: Wearable monitoring of isometric elbow torque is of great interest for rehabilitation monitoring and muscular disease management, yet remains limited by obtrusiveness, drift, and complexity. We present a wearable sensor that measures transient circumferential deformation of the mid-upper arm as a low-complexity bio-signal for joint torque monitoring. Methods: The proposed sensor operates on the basis of magnetoinductive waveguide principles. Participants (n=20) completed three sets of three contractions with sensor replacement between sets. A two-stage analysis extracted contraction-level calibration slopes and characterized variability using a three-level nested mixed effects model. Systematic bias was assessed and leave-one-out cross validation quantified prediction error at the subject, set, and contraction levels. Results: The sensor data had a strong, linear relationship with elbow torque (mean $R^{2}=0.925\pm 0.063$) across the test population. Within-contraction modeling error was $2.5\;Nm$ (5.9%); leave-one-out calibration grouped by set, subject, and population increased error to $5.84\;Nm$ (14.0%), $7.70\;Nm$ (18.4%), and $9.76\;Nm$ (23.4%), respectively, with an opposite trend in calibration burden. No systematic bias was detected across set and contraction order, nor with respect to anthropometric data, supporting the sensor's generalizability. Conclusion: The herein described magnetoinductive waveguide sensor provides a strong, linear predictor of isometric elbow torque with calibration variability emerging primarily due to inter-subject differences that warrant further study. Extension to complex contractions remains critical for future work. Significance: The single-channel, low-complexity, textile-integrated modality enabled by transient circumferential deformation sensing creates a promising path towards monitoring of at-home rehabilitation and muscular disease management.
Stereoelectroencephalography (SEEG)-guided radiofrequency thermocoagulation is the mainstream treatment for drug-resistant epilepsy (DRE), yet non-invasive patient-specific localization of potential epileptogenic zone (EZ) prior to SEEG electrode implantation remains a critical unmet clinical need, hindered by limited automation, suboptimal accuracy, and poor cross-patient generalizability. To address these gaps, we developed an automated non-invasive EZ localization framework that integrates scalp EEG source imaging, high-resolution time-frequency analysis, and deep learning. This multicenter retrospective study included 97 seizure episodes from 37 surgically confirmed DRE patients with Engel Class I post-operative seizure freedom. Three time-frequency spectrograms were generated based on the source-reconstructed signal, and fed into four deep learning architectures (ResNet, VGG, DenseNet, Swin Transformer) for channel-level EZ binary classification with patient-level leave one-out cross-validation to validate personalized localization for unseen patients. The ResNet18-Superlet combination achieved optimal performance (accuracy = 81.15% ± 4.83%, AUC = 84.82% ± 5.27%) in the adult cohort, with robust generalization to pediatric and mixed cohorts, significantly outperforming conventional high-frequency oscillation (HFO)-based methods. This framework enables accurate personalized pre-surgical EZ map ping to optimize SEEG implantation planning, with open-source code available at https://github.com/wyl1994/Source-code-for Patient-Specific-Non-Invasive-Epileptogenic-Zone-Localization to ensure reproducibility and clinical translation.
OBJECTIVE/BACKGROUND:Optically pumped magnetometer-based magnetoencephalography (OPM-MEG) is advancing toward wearable and high-density sensor configurations, posing significant engineering challenges in miniaturization, thermal management, and scalable integration. Here, we present a high-performance wearable OPM-MEG system that addresses these challenges. METHODS:The system was developed through a system-level co-design of the sensor head and electronic control units (ECUs) with a fully automated control workflow. To reduce thermal dissipation and lower scalp temperature, we implemented a thermally optimized suspended vapor-cell module. RESULTS:This design achieves a miniaturized sensor head (12 × 16.5 × 22.5 mm3) and ECU footprint (34 × 28 mm2), with a total per-sensor power consumption of 3.5 W (0.7 W allocated to the sensor head). Crucially, the sensor maintains high sensitivity required for detecting ultraweak brain magnetic fields, exhibiting single-axis sensitivity < 7 fT/$\sqrt$ Hz and dual-axis sensitivity < 10 fT/$\sqrt$ Hz, with a bandwidth of 130 Hz. Inter-sensor crosstalk and intra-sensor cross-axis projection error (CAPE) are suppressed to lower than 2%, addressing the critical challenge of signal interference in high-density sensor arrays. Phantom experiments demonstrate accurate source localization with a mean error of 1 mm, while human recordings provide further validation of stable and high-fidelity performance in wearable OPM-MEG settings. CONCLUSION:Collectively, this work establishes a scalable, instrument-level framework for the design, automated operation, and quantitative end-to-end evaluation of high-density wearable OPM-MEG systems. SIGNIFICANCE:This work provides a scalable foundation for high-density wearable OPM-MEG, enabling mobile brain measurements for brain-computer interfaces, cognitive neuroscience, and future translational clinical neuroimaging applications.
Objective: Event-related EEG activity is widely investigated to characterize brain functions. Traditional analyses rely on heavy pre-processing and strong a priori assumptions, which limit reproducibility and may obscure task-relevant neural activity. This study aims to validate an interpretable convolutional neural network (CNN) capable of highlighting frequency-, spatial-, and temporal-domain EEG signatures in an automatic, data-driven, and end-to-end manner. Methods: We simulated single-trial EEG with imposed spatio-temporal or spectral-spatio-temporal modulations in two paradigms: a visual oddball task and a motor task (200 participants and 200k trials overall). An interpretable CNN was applied to each cognitive task at the single-participant level. CNN-derived spectral, spatial, and temporal signatures were compared with ground-truth signatures known from the simulation by computing localization errors and accuracies. Results: Network features reproduced the modulations imposed in the simulations. The network localized neural signatures with high accuracy: average spectral, spatial, and temporal localization accuracies reached up to 85.3%, 97.8%, and 97.1% across tasks, respectively (top-1 prediction). The corresponding average localization errors were well within established EEG resolution limits (down to 0.95 Hz spectral, 5.7 mm spatial, and 30 ms temporal errors). Conclusion: The interpretable CNN accurately recovered task-relevant EEG signatures across domains, thereby supporting the validity of a CNN-based EEG analysis. Significance: This study provides a ground-truth-based quantitative validation of the multi-domain features learned in interpretable CNNs for EEG analysis, establishing a meaningful reference for trustworthy deep-learning tools that can enhance participant-specific EEG interpretation. These individualized tools could enhance our comprehension of brain functions in both healthy participants and patients.
