High-performance computer-aided drug design is a promising field, in which drug-target affinity (DTA) prediction serves as a core step to reduce R&D costs and improve efficiency. However, current DTA methods suffer from insufficient multi-modal representation and lack interpretability. We propose MAIDTA, an interpretable attention-based multi-modal model integrating sequence and graph dual-modal representations of drugs and proteins. An attention-driven binding region module quantifies drug atom-protein residue interaction strength, and a cross-scale feature fusion module integrates multi-scale features. Experiments on Davis and KIBA show MAIDTA outperforms state-of-the-art models with higher accuracy, robustness, and interpretability, offering new insights for drug discovery.
Patient-specific musculoskeletal models from biplanar radiographs offer a clinically feasible route to estimating spinal loading, but how vertebral landmark error affects force prediction remains unclear. Paired radiograph landmarks reconstruct the three-dimensional spine, making placement error a source of uncertainty in load estimates. We quantified how landmark-placement error affects compression and shear across perturbation magnitude, direction, and spinal region. Using perturbed landmarks and multi-rater annotations, compression maintained excellent ground-truth agreement up to 6.7 mm error. Shear reliability degraded at 2.2 mm, corresponding to 80 N mean absolute error and 26% normalized error. Therefore, these shear estimates require cautious clinical interpretation.
This study combined automated medical image segmentation, hexahedral meshing, and dynamic finite element stance-phase gait simulations with statistical shape modeling and supervised learning for 483 natural knees from the Osteoarthritis Initiative. Ridge regression and recurrent neural networks were trained to predict 38 time-series kinematic, loading, contact, and soft-tissue outputs from anatomic features. The best model, a bidirectional LSTM, achieved an average 1σ-normalized RMSE of 0.45, with inference in seconds. This population-scale framework enables rapid, subject-specific estimates of knee mechanics from imaging alone, supporting future clinical translation.
In contrast to general motor imagery involving large body parts, research on finger motor imagery is very scarce. Due to more refined motor functions, the decoding of finger motor imagery is more arduous and has lower accuracy than that of general motor imagery. In order to improve the decoding accuracy of finger motor imagery, this paper proposes the problem of identifying complementary multi-view decoding features and the problem of electrode channel difference of convolution kernels. A novel multi-branch heterogeneous network (MBHN) consisting of three groups of diversified branches is proposed to effectively extract and exploit complementary features of raw-view, frequency-decomposition-view and wavelet-view. Moreover, the channel adaptive kernel (CAK) module is proposed as a solution to the problem of channel difference of convolution kernels. The experimental results on the public finger motor imagery dataset show that our MBHN model achieves the state-of-the-art decoding accuracy of 58.49%. Additionally, the integration of the auxiliary supervision mechanism and the hybrid loss function is a very effective approach to fully leverage the complementarity of multi-view deep features. Our code is publicly available at https://github.com/ykhdu/MBHN.
Existing dyslexia detection methods typically rely on either EEG or screening tests. This study introduces a multimodal two-stage methodology for dyslexia detection that integrates both EEG signals and screening test data, using spectral features (Shannon entropy and Power Spectral Density) across various frequency bands. The novelty of the proposed approach stems in collecting firsthand paired EEG recordings and screening data from patients, and combining them via two-stage framework, which improves predictive accuracy and yields a comprehensive neurological assessment. Several machine learning models were trained and, upon evaluation, the Random Forest classifier performed best, achieving 98.36% accuracy, 99.2% precision, and 97.63% recall using only EEG data. When both EEG and screening data were combined, performance improved to 98.69% accuracy, 98.90% precision, and 98.90% recall, significantly outperforming existing methodologies, with results validated via statistical tests. Dyslexic individuals exhibited higher delta and theta energy, particularly in the central, parietal, and temporal regions, indicating deficits in phonological processing and working memory. Lower Shannon entropy values in dyslexics suggest increased randomness in neural activity and reduced processing ability. Feature importance analysis identified delta, theta, and gamma energies as the most discriminative features. This study highlights the effectiveness of multimodal data fusion in enhancing dyslexia detection, improving early diagnosis, and facilitating interventions for children aged 5-9 years. These results underscore the potential of advanced spectral feature extraction and robust machine learning in dyslexia diagnostics, setting a benchmark in the field, with future work focusing on validating these results using larger datasets and exploring additional predictive features.
This study develops lightweight Kevlar-polypropylene hybrid composites for prosthetic socket applications using experimental and machine-learning approaches. Laminates containing 0-2 wt.% silica nanoparticles were fabricated and evaluated for mechanical and puncture-energy performance. The 1.5 wt.% silica composite showed the best overall response, reaching a puncture load of 3459.1 N, tensile strength of 225 MPa, flexural strength of 157.24 MPa, interlaminar shear strength of 34 MPa, impact strength of 1813.6 J/m, and net energy absorption of 18.44 J. An artificial neural network achieved R² = 0.983, confirming reliable predictive capability and supporting the composite's suitability for advanced prosthetic applications.
Subject-specific finite element models were developed for 20 primary open-angle glaucoma (POAG) and 20 control eyes using swept-source OCT. The spatial distribution of tensile and compressive strains was computed across the lamina cribrosa. POAG subjects exhibited significantly higher compressive (-0.84 ± 0.15 vs -0.54 ± 0.16) and tensile strains (0.72 ± 0.12 vs 0.54 ± 0.08, p-value < 0.0001). Localized strains within the temporal and superior regions were identified as primary mechanical features distinguishing both among POAG subgroups and between POAG and control subjects. Significant correlation was found between features of strain distribution and visual field mean deviation (p-value < 0.05), defining POAG subgroups that may reflect different POAG mechanical phenotypes.
Fetal arrhythmia is a critical medical condition associated with perinatal morbidity and cardiac complications. This paper proposes an Advanced Detection of Fetal Arrhythmia utilizing Spatial Deep Convolutional Neural Network with GHA-DenseNet (DFA-SDCNN-GHANet). Initially, fetal ECG signals are preprocessed using Adaptive Fast Desensitized Kalman Filter (AFDKF) to remove artefacts, followed by Synthetic Minority Over-sampling Technique (SMOTE) for class balancing. Iterative Local Maximum Synchrosqueezing-Extracting Transform (ILMSET) extracts informative features, which are classified using SDCNN-GHANet optimized by the Black-Winged Kite Algorithm (BWKA). The proposed framework achieved 6.62%, 7.94%, and 9.75% higher F1-score than existing methods.
The cytoskeleton mediates osteocyte responses to fluid shear stress from matrix deformation. Under simulated microgravity, cytoskeleton depolymerizes, while Spectrin-MS provides compensatory structural support. We modeled osteocytes with zero to three crosslinking levels under 1.5 Pa shear. Increased crosslinking raised membrane displacement (+2.60%) and stress (+20.71%), but reduced nuclear displacement (-12.93%) and stress (-10.86%). Cytoskeleton displacement and stress decreased with more crosslinking, whereas Spectrin-MS compensated. Crosslinking alters the intracellular mechanical environment, with Spectrin-MS playing a compensatory role, enhancing understanding of osteocyte mechanosensitivity. .
Kinetic energy non-lethal projectiles (KENLPs) can cause significant thoracic injuries, highlighting the need for computationally efficient tools in impact biomechanics. This study presents the Modified Nonlinear Lobdell Thorax Model (MNLTM), extending the original Lobdell architecture with nonlinear Hunt-Crossley contact and a progressive structural spring. A hybrid global-local strategy combining Latin Hypercube Sampling, Genetic Algorithm, and Sequential Quadratic Programming identifies a single parameter set across three experimental impact cases. The MNLTM reproduces displacement and force responses with improved agreement, achieving 77.0-85.4% displacement and 55.0-97.0% force corridor compliance. The model enables rapid thoracic simulations while reducing reliance on extensive physical testing.
Sleep spindles are transient sigma-band oscillations occurring mainly during NREM Stage N2 and are associated with thalamocortical synchronisation and memory consolidation. Manual spindle scoring remains the clinical gold standard, yet it is labour-intensive and subject to inter-scorer variability. Many automated detectors depend on extensive preprocessing, handcrafted features, or black-box learning models, which may limit interpretability and real-time feasibility.We present an interpretable and computationally efficient spindle-detection framework based on adaptive Kalman filtering (AKF) and innovation variance. First, a proof-of-concept AKF models a single 30 s N2 epoch as a second-order kinematic state-space system and uses the time-varying innovation variance as a direct marker of local nonstationarity. Spindles are detected by a minimal thresholding rule applied to innovation variance, without band-pass filtering, envelope extraction, or time-frequency decomposition. On the pilot epoch, the AKF identifies two spindle events consistent with expert visual annotation, whereas a sigma-band Hilbert-envelope detector and a Martin-type RMS detector capture only a short, high-amplitude segment of the second spindle.Second, we generalise the approach to nine full-night recordings and introduce an AKF-Balanced (AB) detector. AB combines sigma-band filtering with AKF estimation in a three-state (amplitude-velocity-acceleration) model and delineates events using energy-like criteria and innovation-based morphological validation. Compared with two RMS-based detectors, a wavelet/Gabor-like detector, and the innovation-only AKF, AB yields physiologically plausible spindle densities and realistic duration statistics while maintaining a balanced overlap profile against a conservative RMS reference. Overall, model-based adaptive Kalman filtering with innovation variance provides a transparent, real-time-compatible alternative for sleep spindle detection.
Excess heat accumulation and elevated interface pressures cause discomfort and skin complications for transfemoral prosthesis users. This study investigates the thermo-mechanical behavior of silicone liners reinforced with microencapsulated phase change material (PCM) at 5%, 10%, and 15% volume fractions (Si-5PCM, Si-10PCM, Si-15PCM), using a coupled 3D finite element model with multiscale homogenization. Results show the silicone liner reached the highest temperature (∼36°C) and highest coupled contact pressure (183.4 kPa), while Si-15PCM maintained the lowest temperature (∼34.6°C) with smaller pressure and shear-stress increases. Low-fraction PCM liners thus reduce heat-induced mechanical amplification, offering a more stable, skin-friendly prosthetic interface.
Robotic exoskeletons offer significant potential for enhancing human capabilities and supporting medical rehabilitation. However, their control remains challenging due to nonlinear dynamics, parameter uncertainties, and external disturbances. To address these issues, this study proposes an AI-based adaptive control framework integrating a Second-Order Sliding Mode Controller (SOSMC) with a novel Coot Inherited Coati Optimization (COTI-CO) algorithm. The proposed hybrid optimizer combines the global exploration capability of the Coati Optimization Algorithm (COA) with the local exploitation capability of the Coot Optimization Algorithm (COOT) to optimally tune SOSMC gains. MATLAB simulations demonstrate improved tracking accuracy, faster convergence, reduced computational cost, and enhanced robustness under disturbances compared with GWO, COA, COO, and WO.
In artificial intelligence (AI), data dramatically impacts AI models performance, accuracy, and reliability. High-quality data enables models to make better predictions and produce more reliable outcomes. Poor data quality or lack of data can lead to flawed results and cause poor performance and predictions. Sensitivity analysis could play a vital role to generate synthetic dataset from any validated model. Within this work, a sensitivity-based approach is used to generate synthetic data or measurements for healthy and stenosed carotid arteries. Pressure and flow time-series are collected for different levels of stenosis ranging from 30% to 90%, which are created artificially from a validated model of the stenosed carotid arteries. Based on the generated dataset, a machine learning model could be developed for new data to detect the site and level of stenosis in the carotid arteries. In order to validate the applicability of the generated dataset for AI-driven analysis, a proof-of-concept Random Forest classifier was trained to classify different levels of stenoses (mild, moderate, and severe) using a 70%n30% train-test split. The model achieved an overall accuracy of 99.81%, demonstrating strong discriminative capability of the simulated pressure and flow time-series. Although this high performance is obtained under idealized synthetic conditions without measurement noise, the results confirm that sensitivity-based synthetic data generation can effectively support machine learning applications for carotid stenosis detection and severity assessment.
This study proposes a composite control framework for upper-limb rehabilitation exoskeletons integrating fractional-order impedance control, computed-torque compensation, and a momentum-based nonlinear disturbance observer. By combining Jacobian-transpose force mapping, fractional-order impedance shaping, and observer-based compensation, the framework enhances compliant yielding and tracking robustness under non-ideal dynamics. Simulation results show that, under severe spasticity-like impact, interaction force remains within 45 N. Under 30% mass mismatch and strong friction, tracking accuracy improves without reaching the 60 N·m torque limit. Under tremor-like disturbance, 4-8 Hz oscillations are attenuated while low-frequency voluntary motion is preserved, improving compliance, disturbance rejection, and torque smoothness.
Local hyperthermia is an effective treatment for advanced malignant tumors, whose multilayer structure and physiological heterogeneity strongly influence their response to therapy. In this work, a generalized multilayer reaction-diffusion model is developed to study tumor-volume dynamics under localized hyperthermic treatment. The tumor is represented as a set of interacting biological layers, allowing spatial variation in key parameters such as diffusion and proliferation. The model is formulated with appropriate initial, interface and boundary conditions and solved using the finite element method. Simulations are carried out for planar, cylindrical and spherical tumor geometries to examine the effects of heterogeneity, geometry and hyperthermia. The results show a clear reduction in tumor volume under thermal therapy, with stronger effects observed in spherical tumors due to enhanced diffusion. Sensitivity analysis and comparison with published data further support the reliability of the model. Overall, the proposed work provides a simple and flexible approach for analyzing tumor behavior under hyperthermic treatment. This work can support future treatment planning studies.
The anatomical complexity of the mandibular condylar base renders open reduction and internal fixation challenging for fractures in this region. Although the double-straight-plate fixation remains the most widely adopted clinical strategy, its implementation is frequently hindered by technical limitations inherent to conventional operative routes. The biomechanical adequacy of single-geometric-plate fixation has not been conclusively established and warrants rigorous experimental validation. This study developed personalized A-shaped plates with 5-, 6-, and 7-hole configurations, geometrically optimized to adjust the curvature of the anterior arm to accommodate the anatomical contour of the sigmoid notch. Subsequently, the biomechanical efficacy of these personalized plates in stabilizing condylar base fractures was systematically evaluated. Finally, a multi-dimensional quantitative comparison was conducted to assess biomechanical differences between the proposed single-geometric-plate fixation and conventional fixation protocols. The 7-hole-A-shaped-plate configuration demonstrated biomechanical performance equivalent to the double-straight-plate configuration across all parameters. The 6- and 5-hole-A-shaped-plate configuration achieved comparable global stability but exhibited significantly elevated stress concentration in either fixation hardware or adjacent cortical bone. The in silico study furnishes robust biomechanical evidence to support individualized selection of A-shaped geometric fixation for condylar base fractures.
Brain Computer Interfaces promote seamless interaction between individuals with movement limitations and their surrounding environment by transforming electroencephalography signals derived using Motor Imagery. The procedure relies on accurately classifying various MI activities, which requires dependable approaches for EEG signal classification to be continuously improved. In this paper, discrete wavelet transform, and chriplet transform are proposed to enhance the performance of test system with demonstrating critical importance of time-related data and is implemented in visual studio code python. The proposed method holds 91% efficiency, accuracy 94.8% for CBCIC and 93.72% for BCI Competition IV Dataset with response time of 1.03 sec.
Motor imagery (MI)-based brain-computer interfaces (BCIs) decode EEG signals into control commands. However, fine-grained MI decoding within the same limb remains challenging due to highly similar neural patterns. This paper proposes a Contrastive Learning Network based on a Multi-Scale Transformer (CLMT-Net) for fine-grained MI decoding. CLMT-Net integrates multi-scale temporal convolution, FFT-based frequency fusion, spatial convolution, and dual-path Transformer to learn complementary EEG representations. Supervised contrastive learning further improves feature discrimination. On the MI-2 dataset, CLMT-Net achieves an accuracy of 76.13 ± 6.77% with a 95% confidence interval of [73.33, 78.92], demonstrating competitive performance for same-limb MI decoding.
Gallstone disease is a gastrointestinal condition requiring accessible, low-cost screening tools. We developed an interpretable ensemble stacking model for noninvasive gallstone prediction using bioimpedance/laboratory data from 319 individuals (157 patients, 162 controls) with 38 features. Bayesian-optimized SVM, logistic regression, and decision tree models were combined with a deep neural network meta-learner. Ten repeated nested cross-validation yielded 82.3% accuracy and 85.7% AUC, outperforming classical stacking and eight algorithms. SHAP identified CRP, liver enzymes, and HDL cholesterol as key, clinically consistent contributors. However, substantial multicollinearity (VIF >100) caused directional instability for vitamin D and diabetes. Prospective multicenter validation remains necessary before implementation.
