Annual International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE Engineering in Medicine and Biology Society. Annual International Conference最新文献

This study introduces an innovative framework, Ensemble Guide Fine-tuning (EGFit), for Pre-Trained Models, designed to address the challenges of data scarcity in kinase targeted drug discovery. Protein kinases play a pivotal role in cancer, immune diseases, and other complex disorders, making them a critical drug target. Despite over 100,000 recorded kinase inhibitors, only about 75 small-molecule kinase drugs have received FDA approval, underscoring the difficulty of developing kinase drugs. EGFit combines pre-trained large language models, with advanced machine learning techniques, including random forest, support vector machine, multilayer perceptrons, and logistic regression, to iteratively evaluate and refine generated compounds. Under limited data conditions, the framework efficiently explores a vast chemical space, producing biologically relevant and structurally diverse kinase inhibitors. Experimental validation on four kinases (EGFR, MET, PIM1, and CDK5) demonstrates significant improvements in compound similarity to known inhibitors while maintaining compliance with drug-likeness criteria. The iterative feedback mechanism further ensures chemical novelty and biological significance, showcasing the potential of EGFit to optimize compound generation for kinase-specific applications. This framework offers a scalable and effective solution to the challenges of kinase drug discovery, accelerating the development of novel therapeutics and paving the way for broader applications in future studies.

This paper investigates methods for estimating interaction forces during the dynamic process of upper limb elevation. By collecting data from two modalities- electromyographic (EMG) signals of the unilateral forearm and joint angles-along with synchronized interaction force data, a deep learning methodology merging convolutional neural network and long short-term memory network (CNN-LSTM) is adopted to generate a predictive model for characterizing dynamic interactive force, ultimately achieving the task of dynamic force estimation. The comparative analysis of estimation performance using two types of data, EMG signals and EMG-inertial measurement unit (IMU) signals, along with the performance comparison between the CNN-LSTM model and support vector regression (SVR) model for the dynamic force estimation task, demonstrates the advantages of multimodal data and the CNN-LSTM model in facilitating the estimation of dynamic interaction forces in the upper limb.

Traumatic brain injury (TBI) impairs sensorimotor functions, which affect static, dynamic, and reactive balance even in chronic stages. Varying levels of deficits in people with TBI (pwTBI) due to heterogeneous injury pose challenges on therapies. Since qualitative assessment-based conventional therapies are multifactorial, they may not precisely evaluate deficits or provide targeted therapy. However, robotic devices can precisely evaluate deficits and offer customized therapy progression based on deficits. Therefore, the study objective was to investigate the efficacy of a targeted robotic balance training (RBT) in pwTBI using biomechanical and functional outcomes. Data are presented for a small sample of pwTBI who received RBT (TBI-I) and for those who did not (TBI-C). After 10 sessions of training, TBI-I improved in biomechanical (static, dynamic, and reactive balance as well as limits of stability) and functional (community mobility and balance scale) outcomes. These results underscore the preliminary efficacy of RBT in improving balance and postural control in chronic TBI.Clinical Relevance - The data support the efficacy of RBT that can deliver targeted therapy for pwTBI.

Crawling Wave Sonoelastography (CWS) is a quantitative elastography technique that employs two mechanical actuators to generate an interference pattern within the tissue. Ultrasound imaging is then used to capture the resulting wave fields, and the shear wave speed (SWS) is computed to produce an elastography image. In previous studies, different time-frequency techniques have been employed to estimate the SWS, but some limitations, such as lateral artifacts and blurred SWS maps, were reported. In this paper, a novel approach based on the Fourier Synchrosqueezed Transform (FSST) is presented. To assert the veracity of the results, previous datasets in homogeneous and heterogeneous phantoms with vibration frequencies between 200 and 360 Hz have been used. The proposed metrics for comparison were SWS mean value and standard variation, coefficient of variation (CV), Bias, R₂₀₈₀, and, contrast-to-noise ratio (CNR). The new estimator demonstrates marginally superior performance in SWS mean value (at 340 Hz, inclusion: 5.13±0.01 m/s, background: 3.42±0.02 m/s) CV (at 320 Hz, inclusion: 0.11%, background: 0%) and CNR (at 320 Hz, 104.7 dB), and better performance in Bias (at 320 Hz, inclusion: 0.6%, background: 0.84%) and R₂₀₈₀ (at 320 Hz, 0.5 mm) in comparison with previous time-frequency approaches.Clinical relevance- This investigation presents a new Shear Wave Speed estimator for Crawling Waves Sonoelastography approach, which is able to quantify stiffness tissue with great accuracy showing the potential of real-time time application to allow the characterization of tissue elasticity.

Artificial intelligence (AI) has become indispensable in medical image analysis, with models such as convolutional neural networks (CNNs) and Transformer achieving remarkable success in diagnostic imaging. Despite their impressive performance, these models often lack interpretability, limiting their adoption in clinical workflows where understanding disease-specific features is critical for trust.In this study, we propose an explainability framework that enhances interpretability for multi-label disease classification in chest X-ray (CXR) diagnosis by utilizing the U-Net encoder-decoder architecture. The encoder and decoder outputs are concatenated to effectively capture hierarchical features for the classification of 14 observations in the MIMIC-CXR dataset. To further improve interpretability, we apply gradient-weighted class activation mapping (Grad-CAM) across multiple layers, providing detailed insights into the refinement of hierarchical features and the emphasis on disease-specific regions throughout the network. This integration of U-Net with an explainable AI (XAI) framework enhances transparency in the diagnostic process, supporting more informed and trustworthy clinical decision making.Clinical relevance- This study underscores the importance of interpretability in AI-based radiology. By providing clear Grad-CAM visualizations of disease-specific features, clinicians can more confidently validate model predictions and incorporate these insights into their decision-making processes. Through enhanced transparency, our approach not only improves diagnostic performance, but also fosters greater trust in AI tools, paving the way for these models to serve as robust, clinician-friendly decision support systems in routine radiological workflows.

Wearable electrical sensors offer noninvasive, high-fidelity monitoring of organ-level neuromuscular activity. In gastrointestinal applications, electrogastrography (EGG) enables detection of slow-wave (0.05 Hz) gastric myoelectric activity from the skin surface. However, commonly used current electrode systems with individually placed 3M Red Dot electrodes are bulky, prone to electrode placement variability, and unsuitable for long-term or unsupervised clinical use. Here, we present a comparative evaluation of scalable fabrication strategies for a conformable, adhesive-integrated electrode array designed specifically for continuous, high-resolution EGG (HR-EGG). The array is constructed on thin, flexible polyimide substrates with rounded perforations to improve breathability and is paired with a gentle, silicone-based medical adhesive suitable for sensitive skin. This design enables consistent inter-electrode spacing, reduces user burden, and offers significant conformability improvements over rigid commercial multi-electrode systems. It also offers scalability advantages over soft stretchable arrays requiring cleanroom fabrication. Multiple electrode interface strategies-including Ag/AgCl, dry PEDOT:PSS, and conductive hydrogel coatings-are implemented and characterized using electrical impedance spectroscopy. The final patch design is validated through a representative pre- and post-meal recording, showing reliable capture of gastric slow-wave activity. This work supports scalable deployment of HR-EGG in clinical and research settings, expanding access to noninvasive gastrointestinal diagnostics.

Walking on compliant surfaces, such as carpets, grass, and soil, presents a unique challenge, particularly for those relying on prosthetic interventions. Ensuring the safety, stability, and fluidity of movement on these surfaces is paramount to prevent falls and related balance issues in this population. This study presents the first attempt to classify and predict surface compliance in individuals with transtibial lower-limb amputations. By integrating electromyographic (EMG), kinematic, and kinetic data, our system effectively distinguishes user intent across varying surface stiffnesses representing diverse real-world terrains. As we demonstrate the algorithm's success within a clinical population, we achieve up to 83% prediction accuracy, attaining comparable results as in previously tested healthy populations. The suggested framework is a critical component for high-level controllers for advanced prostheses and it holds potential for real-time integration, enabling adaptive adjustments to the prosthetic device in response to both user intent and environmental stimuli.

Ultrasound image segmentation is often limited by the scarcity of annotated datasets, especially in resource-constrained clinical settings. To address this issue, we employ BT-UNet, a self-supervised learning framework that combines Barlow Twins (BT) with the UNet architecture, and aim to enhance segmentation performance in low-data conditions. Unlike previous work that trains BT-UNet exclusively on clinical datasets, our approach explores the benefits of pre-training BT-UNet on musculoskeletal phantom ultrasound images, before fine-tuning it on a small set of annotated clinical images. Our results demonstrate that this strategy significantly improves segmentation performance under limited annotated data. Specifically, with only 5% of the labeled clinical dataset, BT-UNet achieves a Dice score of 0.9311, slightly outperforming the standard UNet's 0.9250. However, at an extreme data scarcity level of 1%, BT-UNet maintains a Dice score of 0.7114, whereas UNet drops to 0.2253. These results highlight the potential of self-supervised pre-training on phantom datasets to address data scarcity challenges in medical imaging. By utilizing unlabeled phantom data for representation learning, BT-UNet enhances segmentation accuracy with minimal clinical annotations, offering a promising solution for real-world medical applications where annotated data is limited.Clinical relevance: This study shows that pre-training a self-supervised learning model on musculoskeletal phantom ultrasound images and fine-tuning it with limited clinical data can significantly improve segmentation accuracy, offering a promising solution to reduce reliance on large annotated datasets.

Time-Up-and-Go (TUG) is a commonly used clinical test to evaluate an individual's gait and frailty state. By combining TUG data with other knowledge, e.g., nutrition and daily habits, informed decisions can be made to delay the progression of or alleviate chronic diseases, such as Parkinson's. Scheduling TUG tests in clinics requires assisted transportation and appointment. With the increasingly overloaded healthcare system, recent advances in e-Health provide an alternative solution. Research studies suggest that it is feasible to perform tests at home and automate gait analysis using intelligent software to classify frailty levels in a remote setting. This allows more frequent monitoring, and clinical appointments are made only to patients at higher risk or those in need. However, conducting the TUG test at home comes with challenges. In this paper, we discuss these challenges, e.g., cluttered environment, and propose solutions. In addition, we investigate whether Body Mass Index (BMI) and gender can affect gait measurement. Our experimental results demonstrate that some machine learning models perform better and the choice of input parameters plays an important role in the classification accuracy. Our experimental results demonstrate that high BMI can be reflected in an individual's TUG, if a robust machine learning model is deployed, while men and women in general show distinct gait measurements. Based on this finding, different thresholds should be defined when making the frail, pre-frail and healthy assessment.

Earables, due to their unobtrusive and lightweight nature, are increasingly being recognized for their potential in estimating digital biomarkers, yet their application in gait analysis (GA) remains limited because comprehensive analytic tools are missing. Existing ear-worn systems have primarily addressed isolated aspects such as gait classification, stride time, or step length estimation, lacking a full end-to-end pipeline. Such pipelines are essential for efficient and automated workflows and real-world applications. This work presents a complete end-to-end GA pipeline for ear-worn accelerometers incorporating multiple algorithms to process raw sensor signals into spatio-temporal parameters. This multi-step approach includes gait sequence detection, event identification, and parameter estimation. We introduce a novel gait sequence detector (GSD) that automatically detects regions of interest in continuous recordings. The integrated spatio-temporal algorithms have already been validated in an isolated setting as part of a previous evaluation study. Using a dataset with three walking speeds and foot-worn IMUs as references, the GSD effectively detects 91 % of gait sequences. The pipeline achieves stride time and SL errors of around 4 % and a gait velocity error of 5.7 %, consistent with prior evaluation for the individual isolated steps. To our knowledge, this is the first end-to-end GA pipeline for earables. Furthermore, the pipeline was released as open-source toolbox (https://github.com/mad-lab-fau/eargait), to facilitate research access and reusability. Our work lays the foundation for automated, continuous, and long-term mobility assessment in home environments using lightweight, unobtrusive earables.