搜索结果：IEEE open journal of engineering in medicine and biology

Objective: Human movement augmentation through supernumerary effectors is an emerging field of research. However, controlling these effectors remains challenging due to issues with agency, control, and synchronizing movements with natural limbs. A promising control strategy for supernumerary effectors involves utilizing electroencephalography (EEG) through motor imagery (MI) functions. In this work, we investigate whether MI activity associated with a supernumerary effector could be reliably differentiated from that of a natural one, thus addressing the concern of concurrency. Twenty subjects were recruited to participate in a two-fold experiment in which they observed movements of natural and supernumerary thumbs, then engaged in MI of the observed movements, conducted in a virtual reality setting. Results: A lightweight deep-learning model that accounts for the temporal, spatial and spectral nature of the EEG data is proposed and called BandFocusNet, achieving an average classification accuracy of 70.9% using the leave-one-subject-out cross validation method. The trustworthiness of the model is examined through explainability analysis, and influential regions-of-interests are cross-validated through event-related-spectral-perturbation (ERSPs) analysis. Explainability results showed the importance of the right and left frontal cortical regions, and ERSPs analysis showed an increase in the delta and theta powers in these regions during the MI of the natural thumb but not during the MI of the supernumerary thumb. Conclusion: Evidence in the literature indicates that such activation is observed during the MI of natural effectors, and its absence could be interpreted as a lack of embodiment of the supernumerary thumb.

Goal: Protein-ligand binding complexes are ubiquitous and essential to life. Protein-ligand binding affinity prediction (PLA) quantifies the binding strength between ligands and proteins, providing crucial insights for discovering and designing potential candidate ligands. While recent advances have been made in predicting protein-ligand complex structures, existing algorithms for interaction and affinity prediction suffer from a sharp decline in performance when handling ligands bound with novel unseen proteins. Methods: We propose IPBind, a geometric deep learning-based computational method, enabling robust predictions by leveraging interatomic potential between complex's bound and unbound status. Results: Experimental results on widely used binding affinity prediction benchmarks demonstrate the effectiveness and universality of IPBind. Meanwhile, it provids atom-level insights into prediction. Conclusions: This work highlight the advantage of leveraging machine learning interatomic potential for predicting protein-ligand binding affinity.

Goal: Accurate detection of malaria parasites using convolutional neural networks (CNNs) relies heavily on the quality of training annotations, yet creating quality annotations is both time-consuming and difficult to scale in high-burden, resource-limited settings. To address this challenge, we propose a method of annotating thin-smear blood images for the semantic segmentation of Plasmodium species, their developmental stages. Using a balanced collection of images from the Malaria Parasite Image Database, we trained identical SegNet models under three matched annotation regimes: expert manual labeling, SegNet-Only, and SegNetOntology where predictions are refined through biomedical ontological reasoning. Model performance was assessed not only for segmentation quality but also for how well each approach captured biologically meaningful information and for its interpretability as judged by clinicians. The proposed method produced results comparable to those achieved by expert annotations and clearly outperformed the baseline SegNet-only model in terms of biological consistency and clinical trustworthiness. The method successfully filtered out 5.7 of invalid AI-generated annotations by identifying semantic contradictions, ensuring the final training dataset adhered strictly to established biological constraints. Clinicians found the outputs from the proposed model nearly as reliable and understandable as those generated from expert annotations. These findings show that embedding formal biomedical knowledge into the annotation process can substantially reduce the cost and effort of creating training data while maintaining diagnostic accuracy and interpretability.

Goal: To quantitatively assess the impact of incorporating radiologist-defined Region of Interest (ROI) information in training deep learning models for thyroid ultrasound image classification and lesion localization. We compared a conventional convolutional neural network (CNN) trained without ROI information, interpreted through Grad-CAM for attention visualization, to Faster R-CNN and YOLOv2 models trained with radiologist-validated ROI masks. We also introduced an adapted mosaic-based composite input, derived from mosaic augmentation but implemented as fixed 1 2 and 2 2 layouts, to improve class balance and spatial diversity in training. Models trained with ROI guidance achieved higher performance in both localization and classification compared to those trained without ROI. The average classification accuracy increased from about 80 in the baseline CNN to around 85 in ROI-guided models that shows an improvement of approximately 5 percentage points. The mean intersection over union between detected and radiologist-defined ROIs increased from approximately 33 to over 70. The adapted mosaic input further stabilized performance across epochs and improved sensitivity while maintaining comparable specificity. Incorporating radiologist-defined ROI information and structured mosaic inputs significantly improves both diagnostic accuracy and localization precision. These results demonstrate that integrating ROI-guided learning with context-preserving composite inputs provides a reproducible framework for developing reliable AI systems in thyroid ultrasonography.

Goal: This study addresses the inherent difficulties in the creation of neuroengineering devices for real-time neural signal processing, a task typically characterized by intricate and technically demanding processes. Beneath the substantial hardware advancements in neurotechnology, there is often rather complex low-level code that poses challenges in terms of development, documentation, and long-term maintenance. Methods: We adopted an alternative strategy centered on Model-Based Design (MBD) to simplify the creation of neuroengineering systems and reduce the entry barriers. MBD offers distinct advantages by streamlining the design workflow, from modelling to implementation, thus facilitating the development of intricate systems. A spike detection algorithm has been implemented on a commercially available system based on a Field-Programmable Gate Array (FPGA) that combines neural probe electronics with configurable integrated circuit. The entire process of data handling and data processing was performed within the Simulink environment, with subsequent generation of hardware description language (HDL) code tailored to the FPGA hardware. Results: The validation was conducted through in vivo experiments involving six animals and demonstrated the capability of our MBD-based real time processing (latency <= 100.37 &#xb5;s) to achieve the same performances of offline spike detection. Conclusions: This methodology can have a significant impact in the development of neuroengineering systems by speeding up the prototyping of various system architectures. We have made all project code files open source, thereby providing free access to fellow scientists interested in the development of neuroengineering systems.

Goal: Cardiovascular disease is the leading cause of death in the USA. Coronary Artery Disease (CAD) in particular is responsible for over 40% of cardiovascular disease deaths. Early detection and treatment are critical in the reduction of deaths associated with CAD. Methods: Sound signatures of CAD vary for individual patients depending on where and how severe the blockage is. We propose the use of the artificial intelligence (AI, specifically the DeepSets architecture) to learn patient-specific acoustic biomarkers which distinguish heart sounds before and after percutaneous coronary intervention (PCI) in 12 human patients. Initially, Matching Pursuit was used to decompose the sound recordings into more granular representations called 'atoms'. Then we used AI to classify whether a group of atoms from a single segment are from before or after PCI. Leveraging the model's learned latent representation, we can then identify groups of atoms which represent CAD-associated sounds within the original recording. Results: Our deep learning approach achieves a test-set classification accuracy of 88.06% using sounds from the full cardiac cycle. The same deep learning architecture achieves 71.43% accuracy using the isolated diastolic window sound segment alone. Conclusions: This preliminary study shows that individualized clusters of atoms represent distinct parts of heart sounds associated with occlusions, and that these clusters differentially change their spectral energy signature after PCI. We believe that using this approach with recordings from individual patients over many time points during disease and treatment progression will allow for a precise, non-invasive monitoring of an individual patient's condition based on unique heart sound characteristics learned using AI.

Goal: In modern high-stress environments, effectively regulating cognitive arousal, through enhancement to boost engagement or inhibition to manage excessive stress, is essential for maintaining mental well-being and optimizing human performance. Hence, this study extends existing state-space models by integrating time-varying parameters and disturbance inputs for enhanced representation of arousal dynamics inferred from skin conductance. Methods: We augmented nominal models with time-varying parameters, then developed a recursive Bayesian estimator for state tracking. Simulation-based validation was performed using skin conductance data from six participants, drawn from an experimental dataset of noninvasive wrist-worn physiological recordings acquired during cognitive stress and relaxation tasks. Adaptive and robust control architectures were designed for closed-loop regulation of latent arousal states. Results: Simulations based on experimental data showed that both controllers outperformed static methods. On average, under inhibitory and excitatory conditions, the adaptive controller achieved average RMSE reductions of 26.9 and 51.6, respectively, while the robust controller achieved reductions of 16.0 and 23.4. In complex multi-step tracking, the adaptive controller reduced average RMSE by 33.7 and control effort by 18.5; similarly, the robust controller reduced RMSE by 32.6 and control effort by 15.1. Conclusion: These findings demonstrate that adaptive and robust control strategies can reliably manage dynamic arousal regulation, offering potential for real-world neuroadaptive systems supporting human performance and well-being.

Objective: Continuous monitoring of blood pressure (BP) is a key parameter for cardiovascular assessment and hemodynamics monitoring. Current noninvasive methods are limited by frequent calibration, motion and environmental artifacts, and delayed response to rapid BP changes. In perioperative and critical-care settings, even short-duration hypotensive episodes and rapid BP lability have been associated with adverse outcomes, motivating technologies with high temporal fidelity. This study introduces a noninvasive blood pressure monitoring technique using superficial temporal artery tonometry (STAT), which employs a biomechanics-based transfer function to improve accuracy, reduce calibration requirements, and detect rapid BP changes in dynamic conditions. Methods: Twenty-nine recording sessions of continuous BP monitoring were collected in human subjects (n [Formula: see text] 10) during rest and during handgrip-induced BP fluctuations. Measurements were recorded simultaneously using the STAT method and compared to a noninvasive reference device (Finapres/Finometer volume-clamp) and Pulse Transit Time (PTT) baseline (derived from timing features) method. Results: Using the Finapres/Finometer as a noninvasive reference, our method achieved a mean absolute difference (MAD) of 4.8 [Formula: see text] 2.2 mmHg during rest and 6.5 [Formula: see text] 3.4 mmHg during handgrips, significantly outperforming PTT, especially under dynamic conditions. Conclusion: BP monitoring with STAT and its biomechanics-based transfer function achieved improved detection of rapid BP fluctuations, and higher accuracy than PTT under dynamic conditions. Significance: STAT with biomechanics-based modeling enables real-time, robust noninvasive BP monitoring, overcoming calibration, motion, and detection limitations of current methods.

Ultrasound imaging is crucial in medical diagnostics, offering real-time visualization of internal anatomical structures. However, accurate automatic segmentation of ultrasound images remains challenging, particularly in scenarios with limited labeled data. In this paper, we propose a semi-supervised learning approach for ultrasound image segmentation, leveraging the statistics of data in unlabeled images to enhance segmentation accuracy. Our method builds upon the encoder-decoder architecture and incorporates innovative semi-supervised learning techniques based on contrastive learning. We have collected ultrasound images from 80 patients and 34 healthy volunteers, focusing on applications in sarcopenia assessment and emergency response scenarios. We demonstrate the effectiveness of our approach through extensive experiments on expert segmentations in this dataset.Our results demonstrate the superior performance of the proposed method across various training data splits (i.e., 1%, 5%, 10%, 20%, 30%, and 100%). While U-NET performed the best with 100% of the training data (i.e., 154 annotated images), the proposed method achieved comparable performance with only 10% of the data (i.e., 16 annotated images). Furthermore, statistical analysis confirmed that our method significantly outperforms existing models, including U-NET, CCT, and UniMatch, in most scenarios (i.e., training set splits). These findings highlight the robustness and efficiency of the proposed method, especially in environments where labeled data is scarce.

Goal: Skull fractures, especially those involving the cranial base and facial regions, present significant diagnostic challenges due to the skull's complex anatomy and subtle radiographic findings. Accurate detection requires repeated and meticulous examination of multiple CT slices, which is a significant cognitive burden, and requires considerable interpretation time. The primary objective of this study is to develop a visualization flattening technique that effectively transforms the curved skull surface into planar representations that enhance fracture features. Methods: A novel visualization process was developed that extracted the cranial surface and subsurface layers from head CT scans and used disk harmonic mapping to generate flattened representations of the lower, upper, occipital, and frontal hemispheres of the skull. The technique was applied to nine cases from the CQ500 dataset, with varying levels of inter-reader agreement, or lack thereof, among the original radiologists who interpreted the dataset. These cases encompass both straightforward and diagnostically challenging fractures that exemplify the advantages of the proposed methodology. Results: The flattened views unwrapped the fractures into continuous, high contrast features, with improved conspicuity compared to the fragmented appearance across multiplanar reconstruction slices. Comparison with existing skull visualization methods, the proposed technique demonstrated high contrast of fractures features, and delineation between emissary veins, with less distortion and high preservation of anatomical continuity. Conclusions: Disk harmonic flattening offers a new approach to skull fracture visualization, providing radiologists and emergency department staff with a valuable addition to the conventional radiological tools, particularly in diagnostically challenging cases.

Goal: Non-compressible torso hemorrhage represents a category of lethal injuries in both civilian and military traumatically injured populations that with proper intervention, training, or technological advancements are survivable. Endovascular localization of active bleeding in the pre-hospital setting can allow faster, less invasive, and more accurate applications of life-saving interventions. In this paper, we report initial in vivo and in silico experimental results to test the feasibility of endovascular localization of hemorrhage. Methods: Endovascular pressure waveforms were acquired on five pigs with an induced aortic injury via a custom intra-aortic catheter instrumented with four pressure sensors. Pressure and velocity data were then simulated on an in silico human aortic model with the same kind of injury. Results: A decrease in pulse pressure across the injury (proximal to distal) reliably indicated the injury location to within a few centimeters. The simulated model showed a similar decrease in pulse pressure as well as an increase in velocity. Conclusions: With additional refinement, localization accuracy may be sufficient for application of a modern covered stent to stop bleeding. The simulated model results indicate relevance for humans and provide guidance for future experiments.

Tinnitus, the perception of sound without an external source, affects many individuals, yet its impact on the brains functional connectome remains underexplored. Traditional functional connectivity (FC) methods, such as Pearson correlation, phase lag index, and coherence, rely on pairwise comparisons between activity of macro-scale brain regions, limiting holistic characterization. We used an approach that estimates the entire connectivity structure by analyzing all time-courses simultaneously, robust even for short recordings and suitable for real-time applications. Using resting-state MEG from tinnitus patients and controls, learned connectomes outperformed correlation-based ones in fingerprinting individuals across test/retest. Group analyses revealed altered FC across multiple frequency bands, impacting default mode, auditory, visual, and salience networks, indicating large-scale reorganization. Tinnitus exhibited highly individualized whole-brain FC profiles, highlighting the importance of individual variability and paving the way for personalized models to optimize patient-specific interventions.

Goal: To develop a compact, real-time microfluidic spectroscopy system capable of simultaneously measuring the concentrations of multiple solutes flowing together through a single fluid pathway with high temporal resolution. Methods: The measurement system integrates a Z-flow cell and dual-wavelength LED light sources with a compact spectrophotometer. The experimental setup consisted of two clinical infusion pumps delivering distinct marker dyes through a common fluid pathway composed of a clinical manifold and a single lumen of a clinical intravascular catheter, while a third pump delivered an inert carrier fluid. Concentration measurements of the mixed dyes were performed at high-frequency sampling intervals, with dynamic pump rate adjustments to evaluate the system's ability to detect real-time changes in solute concentration. A MATLAB-based control application enabled automated data acquisition, processing, and system control to enhance experimental efficiency. Results: The system accurately measured solute concentrations, capturing temporal variations with high precision. It demonstrated high reproducibility with a standard error of the mean no larger than [Formula: see text] for Erioglaucine and [Formula: see text] for Tartrazine at steady state, and high accuracy with a maximum deviation of [Formula: see text] for Erioglaucine and [Formula: see text] for Tartrazine from the expected steady-state concentrations. Conclusions: This system enables continuous, real-time monitoring of multiple solutes in dynamic flow conditions, offering a portable solution with high sensitivity to temporal concentration changes-advancing beyond traditional static fluid measurement methods.

Goal: Electroencephalogram-based brain-computer interfaces (EEG BCIs) have broad applications in neurorehabilitation, clinical assessment, and assistive technologies. However, their practical deployment is severely limited by subject-specific calibration, which requires time-consuming data collection and model retraining for each user, significantly reducing usability. This reliance on calibration arises from the conventional "one-model-fits-all" strategy: "relying on a single general model to handle all data complexity like subject variability. When its limited generalization falls short, time must be spent on calibration to adapt the model." Methods: To address this limitation, we propose a trade-space-for-time strategy for calibration-free EEG decoding: "Instead of adapting one model to every user, we maintain a pool of compact models, including a general model and multiple biased models, where each biased model specializes in decoding a specific type of subject pattern. For a new input, the system automatically selects the most suitable model based on data characteristics, enabling instant adaptation without retraining." Compact deep learning models make this design feasible by allowing fast switching and low storage cost, which would be impractical with large-scale architectures. Results: Experiments on multiple public EEG datasets show that the proposed strategy achieves performance comparable to within-subject decoding: slightly higher in one dataset (0.7672 vs. 0.7601), nearly identical in another (0.7568 vs. 0.7572), and marginally lower in a third (0.8804 vs. 0.8888). Conclusions: These results demonstrate that our approach effectively eliminates calibration while preserving accuracy, providing a practical and scalable alternative for EEG BCIs. The framework also has potential applications in other neuroimaging modalities such as fMRI and fNIRS.

Goal: To develop a high-performance and robust solution for neonatal sleep staging that incorporates spatial topological information and functional connectivity of the brain, which are often overlooked in existing approaches. Methods: We propose MVBNSleepNet, a multi-view brain network-based convolutional neural network. The framework integrates a multi-view brain network (MVBN) to characterize brain functional connectivity from linear temporal correlation, information-theoretic, and phase-dynamics perspectives, providing comprehensive spatial topological information. A masking mechanism is employed to enhance model robustness by simulating random dropout or low-quality signal conditions. Additionally, an attention mechanism focuses on key regions of the brain network and reveals structural brain connectivity during sleep, while a CNN module extracts spatial features from brain networks and classifies them into specific sleep stages. The model was validated on a clinical dataset of 64 neonatal EEG recordings using a leave-one-subject-out validation strategy. Results: MVBNSleepNet achieved an accuracy of 83.9% in the two-stage sleep task (sleep and wakefulness) and 76.4% in the three-stage task (active sleep, quiet sleep, and wakefulness), outperforming state-of-the-art methods. Conclusions: The proposed MVBNSleepNet provides a robust and accurate solution for neonatal sleep staging and offers valuable insights into the functional connectivity of the early neural system.

Goal: Emerging evidence in diverse tumor types establishes a link between lymphatic dissemination and collective tumor cell invasion. To simulate the biomechanical features of the tumor-lymphatic microenvironment, we developed a 3D tumor-lymphatic architecture biomimetic (T-LAB) platform. Methods: Mathematical and computational fluid dynamics modeling were used to determine the fluid flow, oscillatory flow-induced shear stress, and system pressure in the 3D-printed macrofluidics platform. Results: Various human breast cancer cell lines and human dermal lymphatic endothelial cells (HDLEC) were seeded on a matrix in the T-LAB and imaged for up to 96 h to assess cell morphology, viability, migration, and invasion. Co-culture of inflammatory breast cancer cells with HDLEC in the T-LAB, determined to simulate the fluidic properties of the tumor lymphatic microenvironment, demonstrated tumor cell clusters/emboli formation and collective invasion similar to the clinicopathological features observed in patients. Conclusions: The 3D T-LAB model developed here can be used to culture any type of tumor cell to study topographical features that impact tumor-lymphatic interface, collective invasion, and lymphatic dissemination.

Objective: This study investigates the neurodynamics of motor imagery speed decoding using deep learning. Methods: The EEGConformer model was employed to analyze EEG signals and decode different speeds of imagined movements. Explainable artificial intelligence techniques were used to identify the temporal and spatial patterns within the EEG data related to imagined speeds, focusing on the role of specific frequency bands and cortical regions. Results: The model successfully decoded and extracted EEG patterns associated with different motor imagery speeds; however, the classification accuracy was limited and high only for a few participants. The analysis highlighted the importance of alpha and beta oscillations and identified key cortical areas, including the frontal, motor, and occipital cortices, in speed decoding. Additionally, repeated motor imagery elicited steady-state movement-related potentials at the fundamental frequency, with the strongest responses observed at the second harmonic. Conclusions: Motor imagery speed is decodable, though classification performance remains limited. The results highlight the involvement of specific frequency bands and cortical regions, as well as steady-state responses, in encoding MI speed.

Background: Post-traumatic stress disorder (PTSD) is a psychophysiological condition caused by traumatic experiences. Its diagnosis typically relies on subjective tools like clinical interviews and self-reports. Objectives: This scoping review analyzes computational methods using EEG signal processing for PTSD diagnosis, differentiation, and therapy. It provides a comprehensive overview of the entire EEG analysis pipeline, from acquisition to statistical and machine learning techniques for PTSD diagnosis. Methods: Using the PRISMA-ScR protocol, studies published between 2013 and 2024 were reviewed from databases including Scopus, Web of Science, and PubMed. A total of 73 studies were analyzed: 52 on diagnosis, 8 on differentiation, and 15 on therapy. Results: EEG Bands and Event-Related Potentials (ERP) were the dominant techniques. The Alpha band demonstrated strong performance in diagnosis and therapy. LPP ERP was most effective for diagnosis, and P300 for differentiation. Supervised SVM models achieved the highest accuracy in diagnosis (ACC = 0.997), differentiation (ACC = 0.841), and psychotherapy (ACC = 0.78). Random Forest multimodal models integrating EEG with other modalities (e.g., ECG, GSR, Speech) achieved ACC = 0.993. Unsupervised approach is employed to cluster patients to identify PTSD subtypes or to differentiate PTSD from other mental disorders. Veterans and combatants were the primary study population, and only three studies reported open datasets. Conclusions: EEG-based methods hold promise as objective tools for PTSD diagnosis and therapy. The review identified limitations in the use of ERP, sleep characterization and full-band EEG. Broader datasets representing diverse populations are essential to mitigate bias and facilitate robust inter-model comparisons. Future research should focus on deep learning, adaptive signal decomposition, and multimodal approaches.

Deep learning-based Generative Models have the potential to convert low-resolution CT images into high-resolution counterparts without long acquisition times and increased radiation exposure in thin-slice CT imaging. However, procuring appropriate training data for these Super-Resolution (SR) models is challenging. Previous SR research has simulated thick-slice CT images from thin-slice CT images to create training pairs. However, these methods either rely on simplistic interpolation techniques that lack realism or on sinogram reconstruction, which requires the release of raw data and complex reconstruction algorithms. Thus, we introduce a simple yet realistic method to generate thick CT images from thin-slice CT images, facilitating the creation of training pairs for SR algorithms. The training pairs produced by our method closely resemble real data distributions (PSNR = 49.74 vs. 40.66, p [Formula: see text] 0.05). A multivariate Cox regression analysis involving thick slice CT images with lung fibrosis revealed that only the radiomics features extracted using our method demonstrated a significant correlation with mortality (HR = 1.19 and HR = 1.14, p [Formula: see text] 0.005). This paper represents the first to identify and address the challenge of generating appropriate paired training data for Deep Learning-based CT SR models, which enhances the efficacy and applicability of SR models in real-world scenarios.