Computer Methods in Biomechanics and Biomedical Engineering最新文献

The timely prediction of preterm labor plays a crucial role in improving neonatal survival, as well as in providing professional care for the infant and mother. The Electrohysterography (EHG) signals have higher sensitivity, which provides a promising avenue for analyzing preterm labor contractions. The conventional preterm labor prediction approaches are highly vulnerable to various limitations regarding higher correlations, a lack of robustness, as well as lower sensitivity to minor variations. The research proposes a multi-head attention-enabled distributed neural network with a bidirectional long short-term memory (multi-DNBiTM) framework that aims to overcome the challenges of conventional approaches by exhibiting accurate predictions. The research employs Root mean energy-entropy deep features (RMEn2D) that are beneficial in analyzing the variations of each frequency subbands, providing a higher representation of minor uterine contractions. The multi-head attention categorizes the feature maps into diverse heads, which facilitates the learning of intrinsic patterns. Moreover, multi-level training improves prediction accuracy through analyzing the signals at different levels and allows to learn fine details. The efficiency of the multi-DNBiTM model is evaluated with the prevailing approaches, which reveal better outcomes by acquiring 96.93% accuracy, 98.45% sensitivity, and 98.19% specificity.

EEG-based emotion recognition has attracted considerable research interest due to its non-invasive characteristics and superior temporal resolution. Challenges still remain in modeling global dependencies, cross-channel interactions, and cross-domain generalization. This study proposes a novel model called DAVG-ViT (a Domain-Adaptive Vision Transformer with GA-BiLSTM for EEG Emotion Recognition). It enhances global modeling using Adapters, captures cross-channel relationships and bidirectional temporal dynamics via Group Attention and BiLSTM, and reduces distribution shift through a gradient reversal layer-based adversarial domain adaptation mechanism. Experiments on SEED, SEED-IV, and DEAP show superior performance, achieving accuracies of 99.37% on SEED, 95.91% on SEED-IV, and 98.95% for arousal and 98.47% for valence on DEAP.

This study introduces a generative machine learning-based framework for the reverse design of puncture needle deflection curves, aiming to address the challenge of precise navigation of puncture needles in soft tissues. The framework consists of two primary components: reverse prediction and forward validation, which work together to derive optimal needle tip design parameters from the target deflection curve. A needle-tissue interaction physics model is developed based on Euler-Bernoulli beam theory, and 10,000 datasets are generated through a parametric sampling method. Experimental results demonstrate that the curve classifier achieves an F1 score ranging from 0.945 to 0.969, while the deflection curve regressor shows an R² value of 0.9981 with an average error of only 0.0008 mm. The average relative error of the predicted design parameters is kept within 5%. Furthermore, an innovative ensemble learning strategy is employed, which leads to a further reduction in the prediction error by 4.3%. In cases involving high-deflection curves, the optimal design (d = 1.01 mm, L = 113.84 mm, θ = 39.08°) was validated via finite element analysis and showed excellent agreement with the target trajectory, with an NRMSE value of 1.15903E - 4 and deviation controlled within ±0.3 mm. This study not only provides an efficient solution for the precise design of medical needle tips but also establishes a new research paradigm for the intelligent design of medical devices.

Brain-computer interfacing facilitates usage of medical devices such as Electroencephalograms to study brain activities using signal processing techniques. Hand movements are motor activities which cause signature electrical signals in the electroencephalogram recordings. Signal processing and machine learning can be used to remove artefact contamination, amplify subtle features associated with hand movement and classify them. This paper experiments to utilize mathematical models to extract features and classify hand movement from electroencephalogram data up to 98% accuracy based on tests performed on an open-sourced dataset. The study, after further tests, can be used to build prosthetic limbs and mind-controlled robotic arms.

Biodegradable poly-lactic acid (PLA) stents reduce the risk of restenosis and late thrombosis, but their clinical adoption is limited by the trade-off between radial strength and deliverability. To address this, a novel auxetic-structured PLA (ASP) coronary stent is proposed. Finite element analysis was conducted to evaluate its performance against commercial Absorb and Fantom stents. The ASP stent achieved a 12.6% higher radial strength than the Fantom stent, along with -31.3% axial foreshortening, as well as reduced radial recoil and plastic strain. This improvements are attributed to the synergistic combination of PLA's properties and auxetic structure, which balances stability with deliverability.

Alternative splicing (AS) generates diverse mRNA isoforms and plays a key role in cancer. Using Drosophila homologous genes, this study identified prognostic AS events across pan-cancer via Cox regression and constructed robust prognostic signatures validated by survival and ROC analyses. AS was shown to significantly influence the tumor immune microenvironment and immunotherapy response. COL1A1 emerged as a strong pan-cancer prognostic factor. Splicing factor regulatory networks and Mendelian randomization analyses further revealed genes associated with tumor development, highlighting AS as a valuable source of prognostic biomarkers and therapeutic targets.

This paper introduces a multi-scale feature fusion framework for EEG-based motor imagery (MI) classification, designed to leverage the spectral-temporal-spatial structure of EEG data, its nonlinear intrinsic characteristics, and convolutional features. Several proposed feature fusion models surpass current state-of-the-art classification systems for MI tasks. A support vector machine (SVM) model achieves an accuracy of 86.92% on the BCIC-IV-2a dataset. To mitigate redundancy, the proposed models incorporate dimensionality reduction via factor analysis (FA) and channel selection using common spatial pattern (CSP). Selecting 12 channels yields superior classification performance compared to using all 22or only 8 selected channels, achieving an accuracy of 88.17%.

To address the limitation that prevailing cross-domain Parkinson's disease (PD) speech recognition methods mitigate small-sample issues via distribution matching while disregarding substantial feature overlap, we introduce the cross-domain Fisher criterion (CFC). CFC reformulates inter- and intra-class scatter matrices to align with classifier properties: the former enhances separation among heterogeneous target-domain samples, whereas the latter compactly aggregates homologous cross-domain samples around the target centroid and suppresses inter-domain disparities. Numerous experimental results demonstrate that CFC is an effective, efficient, and robust method for PD speech recognition, offering a promising approach for integration into PD speech-based remote rehabilitation and monitoring systems..

Accurate prediction of gestational diabetes mellitus (GDM) is critical for improving maternal and fetal outcomes. This study develops a Transformer-based multimodal fusion model that integrates tabular clinical features and image-encoded electronic health records (EHRs), aiming for accurate end-to-end classification of GDM. Preprocessed EHRs were transformed into grayscale, RGB, and heatmap, with visual features were extracted by a Vision Transformer and tabular features by an MLP. A modality-aware attention mechanism enhances cross-modal fusion. Evaluated on two public datasets, performance gains over the strongest single-modality models reached 3.95% and 0.38% in accuracy.

Renal artery stenosis (RAS) is a major cause of secondary hypertension, requiring accurate hemodynamic evaluation for clinical intervention. This study presents a deep learning framework integrating Mamba-based state-space modeling (SSM) with hierarchical point cloud processing for real-time hemodynamic prediction. A computational dataset was generated from three-dimensional renal artery models using Bessel-curve reconstruction and computational fluid dynamics (CFD) simulations. By combining PointNet++ with Mamba's selective mechanisms, the model effectively captures hemodynamic metrics while preserving local vascular features. The method provides real-time renal hemodynamic predictions with computational efficiency improved by several orders of magnitude while preserving accuracy comparable to CFD.