In the current era of precision psychiatry, the diagnosis of psychiatric illnesses, particularly schizophrenia, remains challenging and complex task for clinician. The objective of this study is to bring an help tool in this diagnosis using biomarkers such as electroencephalograms (EEGs) and approaches based on machine learning (ML). In our work, we propose to quantify functional connectivity in schizophrenia using several methods (Phase Lag Index, Weighted Phase Lag Index, Squared Coherence and Shannon Entropy), to extract connectivity graphs from EEG signals and to compute topological features such as clustering coefficient, number of triangles, strength, and eigenvector centrality. These features will be used for classification using ML approaches. The main contribution of our approach is the comparison of several connectivity methods based on the computing time, the accuracy, recall, the classifier algorithm, and the train set size. The choice of the threshold for the connectivity methods is also discussed. The results obtained with around 17 classification model for each connectivity method show that Shannon entropy provides better results with an accuracy of around 98%, a sensitivity and specificity of 96%, and a computing time in the order of a few seconds. Our results show also that it is possible to diagnosis the schizophrenia using EEG as biomarker if the classification pipeline is well designed.
Surface electromyography (sEMG) has become a key sensing modality for hand gesture recognition in rehabilitation, prosthetic control, and human–computer interaction. However, its sparse activation patterns and sensitivity to subject- and recording-related variability pose significant challenges, particularly for fine-grained gestures involving subtle finger and wrist motions. To address these limitations, we propose an attention-enhanced multimodal fusion framework that integrates raw sEMG and accelerometer (ACC) signals. The proposed network employs narrow-kernel temporal convolutions in dedicated modality-specific streams to preserve fine-grained neuromuscular patterns, followed by a dual-attention fusion module combining multi-head self-attention (MHSA) and gated attention to model interactions between sEMG and ACC embeddings and improve gesture separability. Experiments on the NinaPro DB2 and DB7 databases show that the proposed fusion model outperforms both the sEMG-only and ACC-only models. The feature-level fusion model achieves accuracies of 96.42% on DB2 and 97.03% on DB7 without data augmentation. After applying the selected combined augmentation strategy based on time warping and scaling, the accuracy further increases to 97.56% and 98.19%, respectively. A decision-level fusion variant also achieves comparable performance while keeping the sEMG and ACC classifiers independent. Detailed gesture-level analysis further shows that the largest gains are obtained for low-amplitude and highly similar gestures, highlighting the value of multimodal fusion for improving the separability of subtle hand movements.
Background The intervertebral flexion-extension kinematics of the neck are widely studied, but most of the range of motion analyses do not quantify the neutral neck position. Therefore, the influence that the neutral position might have on the kinematics is usually ignored. To improve the understanding and the prediction of intervertebral motion, this study was designed to measure the intervertebral range of motion with 3D geometries obtained from biplanar X-rays, and to explore the impact of the neutral posture on maximal flexion and extension. Methods The subject-specific 3D geometries of the cervical spines (C1-C7) of thirteen male and female volunteers were studied in full flexion, neutral position, and full extension. Biplanar X-ray images were obtained with an EOS low-dose system. The intervertebral values were measured as the angles between the transversal vertebral body midplanes for each vertebral level. Results The intervertebral neutral position correlated with C3-C7 full flexion and full extension. The cervical levels exhibiting greater extension in the neutral position achieved a lower range of motion in full extension (P<0.001). Similarly, higher flexion in the neutral position resulted in a reduced intervertebral range of motion in full flexion (P=0.003). Conclusion Variations in the neutral position may affect full flexion and extension range of motion. We believe it is the first time that these linear relationships considering the neutral neck position are presented in the literature. Considering the neutral position can help more accurately predict kinematics at the end range of the motion.
Objectives Neonatal seizures are critical neurological emergencies requiring prompt diagnosis. Current deep learning methods for electroencephalogram (EEG) analysis often neglect the spatial topology between electrodes. This study aims to develop a novel, lightweight, multi-modal convolutional neural network to address this limitation and improve detection accuracy. Material and methods We propose a multimodal convolutional neural network integrating wavelet transform and attention mechanisms (MWAM-CNN). The model introduces three key innovations: (1) a two-dimensional spatial-topographic encoding (2D-STE) strategy to generate a five-dimensional tensor preserving spatial information for 3D convolutions;(2) a learnable wavelet transform convolution (WTConv) module for multi-scale feature extraction;(3) a four-path parallel architecture with a 3D Convolutional Block Attention Module (CBAM-3D) for adaptive feature fusion. Results On the Helsinki dataset, balanced using the SMOTE method, the MWAM-CNN model achieved an accuracy of 97.2%, a sensitivity of 98.5%, a specificity of 95.4%, and an AUC of 99.7%. This high performance was achieved with a lightweight design of only 10,110 parameters, demonstrating its efficiency and effectiveness. Conclusion The proposed MWAM-CNN provides a robust and efficient solution for neonatal seizure detection. By effectively integrating spatial, temporal, and spectral features through its innovative architecture, the model significantly enhances detection performance while maintaining a lightweight structure suitable for clinical applications.
Background: Measurements of body-composition changes are important for assessing patients' nutritional status and the progression of diseases such as sarcopenia, cachexia, atrophy, and cancer. Ultrasound imaging is the preferred technique for direct body-composition measurements owing to its accessibility, ease of use, and non-invasiveness. Typically, ultrasound imaging techniques use echogenicity changes to identify the interface between different tissue layers, resulting in time-consuming measurements with accuracies that depend on the ultrasound image quality. Radio frequency (RF) signals obtained directly from an ultrasound system ensure the reproducibility of measurements. Purpose: This study proposes a two-dimensional ultrasound signal-processing technique for measuring body composition. Basic procedure: Backscattered RF signals were acquired from the forearms of human subjects using an ultrasound system. The interfaces of subcutaneous fat, muscle, and bone were identified using the proposed signal processing technique for RF signals and ultrasound imaging. The longitudinal-sectional areas of the subcutaneous fat and muscle measured using ultrasound signals were validated by comparison with those measured using ultrasound imaging. Main findings: The results of the proposed technique and ultrasound imaging exhibited strong linear correlation. The area correlations between the proposed technique and ultrasound imaging were 0.937 and 0.991 for subcutaneous fat and muscle, respectively. Conclusions: The proposed technique demonstrates feasibility of clinical applications by performing an in vivo human study. (c) 2026 AGBM. Published by Elsevier Masson SAS. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
Context: Surgical work-related musculoskeletal disorders (WRMSDs) has mostly been investigated using qualitative observational assessments and questionnaires, as well as quantitative marker-based motion analysis. The main limitations of these methods are the subjectivity of questionnaires and the impracticality of markers in operating room. This study investigates the feasibility of predicting ergonomic level and surgical expertise level using quantitative postural metrics. Material and method: Sixty orthopedic surgeons with three expertise levels were recorded without markers using two cameras, frontal and sagittal views. Human pose was extracted from the 120 videos using OpenPose combined with a custom post-processing pipeline. The ergonomic level was assessed by two expert surgeons using a 5-point Likert scale. A large set of quantitative postural metrics, designed with expert surgeons, was computed and used to predict both ergonomic level and surgical expertise level using Random Forest and XGBoost Tree. Results: The best performance for surgical expertise prediction was achieved using Random Forest model, with a balanced accuracy of 48%. For ergonomic level prediction, the lowest error was obtained with the Random Forest model, with a RMSE of 0.77. The most informative postural metrics for predictions were primarily related to upper-limb joints.
Objective: The integration of multi-omics data to uncover the biological mechanisms of human diseases remains a significant challenge in bioinformatics. While deep learning (DL) has emerged as a powerful tool for this task, current methods often fail to model the complex correlations among features and samples, limiting both predictive performance and interpretability. Methods: To address this, we propose MOFRCDLANet (Multi-Omics Feature Reordering Correlations Deep Attention Network), a novel framework for predicting tumour recurrence and identifying biomarkers. Our model introduces a feature reordering strategy to prioritise prognostically relevant features. It then employs a self-attention module coupled with Maximum Mean Discrepancy (MMD) and contrastive regularisation to learn robust latent representations that capture cross-sample relationships and align feature distributions across omics types. Finally, an attribution-based method identifies key biomarkers, providing biological insight into the model's predictions. Results: Extensive experiments on ten TCGA cancer datasets demonstrate that MOFRCDLANet outperforms state-of-the-art methods across key metrics, including accuracy and AUC. The top genes identified by the model were biologically validated through enrichment analyses (KEGG and GO), confirming their relevance to cancer pathways and reinforcing the framework's efficacy. Conclusion: MOFRCDLANet provides a robust, interpretable solution for multi-omics integration, advancing both the predictive accuracy and mechanistic understanding of cancer progression. This work offers a valuable tool for precision oncology, enabling improved prognostic stratification and biomarker discovery. (c) 2026 AGBM. Published by Elsevier Masson SAS. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
Background: Objective assessment of lower-limb muscular strength typically requires maximal-effort testing or laboratory-based equipment, limiting practicality in routine exercise and rehabilitation settings. Wearable exoskeletons provide controlled resistance and quantitative biomechanical monitoring, offering a potential platform for structured strength evaluation. This study proposes an interpretable framework for estimating lower-limb strength using joint torque and surface electromyography (sEMG) metrics obtained during exoskeleton-assisted exercise. Methods: Thirty healthy adults completed conventional strength assessments to construct a composite reference index. Participants then performed guided squat, knee-up, and lunge exercises using a hip-joint exoskeleton. Torque and sEMG signals were used to derive execution-based performance metrics. Decision-tree models were developed for three-level strength classification, and regression models were constructed for quantitative strength estimation. Results: Metrics reflecting exercise execution consistency, particularly guided exercise pace, emerged as the most influential indicators of strength. The expert rule achieved precision up to 0.95 for strength classification, and regression models showed strong association with the composite strength index (maximum r = 0.86, p < 0.001 ). Conclusion: The proposed framework enables interpretable classification and quantitative estimation of lower-limb strength using standardized wearable exercise. Integration of torque and sEMG-derived metrics supports practical and data-driven strength assessment without reliance on maximal-effort testing. (c) 2026 AGBM. Published by Elsevier Masson SAS. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
Background: The metabolic cost of walking reflects the energy required to move the body over a given distance. Standard measurement methods require prolonged exertion from participants, which is not always feasible. Estimating instantaneous metabolic cost is valuable for real-time control of assistive devices such as exoskeletons or prostheses. This can be achieved by computing mechanical work and applying separate efficiencies to positive and negative work. Objective: This study aimed to compute and compare positive and negative work efficiencies using four different methods. A second objective was to evaluate how well metabolic cost, estimated from joint mechanical work weighted by these efficiencies, matched measured values. Method: Eleven participants walked on slopes of +/- 24%, +/- 12%, +/- 8% and level. The +/- 24% slopes were used to represent conditions of predominantly positive (ascent) and negative (descent) work. Mechanical work and efficiencies were calculated using four methods: potential energy (PE), combined limb (CLM), individual limb (ILM), and summed joint (& sum;joints). Metabolic costs estimated from summed joint work, adjusted by the efficiency pairs, were then compared to measured metabolic costs on intermediate slopes. Results: Results showed that efficiencies - particularly for negative work - depended on the mechanical work calculation method (negative efficiency range: -0.91 to -1.08). Despite this, all methods showed strong correlations with measured metabolic cost (r >= 0.97). Moreover, positive efficiency had a greater influence on the estimated metabolic cost than negative efficiency. Discussion: Efficiencies derived from the CLM, ILM, and & sum;joints methods more accurately captured changes between conditions than those obtained from the PE method, suggesting these methods are preferable. Future studies should assess their relevance in other contexts, such as load carriage or speed variations. (c) 2026 AGBM. Published by Elsevier Masson SAS. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Background and Objective: Major Depressive Disorder (MDD) affects a wide range of populations and causes significant harm to individuals and society. Hence, early recognition of MDD is crucial. MDD recognition using wearable electroencephalographic (EEG) devices has gained significant attention, with reliable and effective classification algorithms central to its success. Methods: Herein, an end-to-end framework named an end-to-end shuttle neural network (ESNN), is proposed for efficient recognition of depression on multichannel EEG signals. The ESNN comprises three parts: i) a multiscale saliency-encoded spectrogram that effectively captures time-frequency information from multichannel EEG signals; ii) TSUnet, a two-stream temporal spectrogram U-Net incorporating the crossmodule attention to redistribute feature weights and enhance critical information; and iii) a crosschannel-wise block to integrate time-frequency features from the two-stream network. Results: Two public EEG datasets [the Hospital Universiti Sains Malaysia (HUSM) and MODMA)] and one private EEG dataset [Zhongda Hospital, Southeast University (ZHSU)] were used to confirm the model's performance. The leave-one-subject-out validation experiment was conducted to ensure subject independence. Our proposed ESNN achieved accuracies of 98.70% and 86.36% on HUSM and MODMA datasets, respectively. On ZHSU dataset, the framework remarkably performed with 83.85% accuracy. Conclusion: The results verified that different scale features could be adequately captured by branch processing and fusion of time-frequency information. Ablation experiments also suggested that the proposed crossmodule attention and channel-wise block effectively focused significant information, suggesting that this model could potentially recognize depression in a real-world scenario. Our model exhibits the potential for application as a clinical decision support tool. By assisting physicians in diagnosis, it contributes to the conservation of healthcare resources. The code is provided in: https:// github.com/zf703/ESNN. (c) 2026 AGBM. Published by Elsevier Masson SAS. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
Background and objective: Sprinting-induced fatigue significantly compromises neuromuscular performance and elevates the risk of injury. Accurate monitoring of muscle fatigue is essential for designing personalized training and rehabilitation strategies. This study presents a novel method that integrates hybrid feature extraction with interpretable machine learning to assess fatigue during dynamic contraction. Methods: In this work, time-frequency methods, namely Stockwell Transform (S-transform), B-Distribution (BD), and Extended Modified B-Distribution (EMBD), were applied to distinguish dynamic muscle fatigue states. Surface electromyography (sEMG) signals were recorded from the lower limb muscles of 14 healthy collegiate athletes during sprinting. The non-fatigue, fatigue progression, and fatigue segments of the signals were preprocessed and analyzed using these methods. From each method, thirteen features were extracted, and prominent features were selected using Genetic Algorithm (GA) and Principal Component Analysis (PCA). Classification of fatigue states was performed using four machine learning algorithms: Decision Tree, Support Vector Machine (SVM), Random Forest, and Artificial Neural Network (ANN). Furthermore, spectral features such as mean frequency and median frequency were analyzed to compare fatigue across different muscles. Results: The results demonstrate that fatigue is characterized by a progressive decline in median frequency (MDF) and mean frequency (MNF) over time. The gastrocnemius lateral head exhibited the steepest decrease in both MDF and MNF, indicating a higher susceptibility to fatigue during dynamic contractions, followed by the gastrocnemius medialis head. Comparative analysis of spectral features across sEMG segments revealed that the transition from fatigue progression to established fatigue occurred more rapidly than the shift from non-fatigue to fatigue progression. Classifier performance evaluation showed that the Random Forest model achieved the highest accuracy of 96.62% using features selected by the Genetic Algorithm (GA), outperforming models trained on Principal Component Analysis (PCA)-selected features (90.79%) and all features combined (92.83%). In contrast, the Support Vector Machine (SVM) classifier recorded the lowest accuracy at 66%. Conclusions: The proposed method effectively detects dynamic muscle fatigue and shows strong potential for integration into real-time fatigue monitoring for wearable systems. (c) 2026 AGBM. Published by Elsevier Masson SAS. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
Objectives: Breast cancer remains one of the most prevalent and life-threatening malignancies worldwide, wherein accurate and early diagnosis plays a pivotal role in improving patient outcomes. The screening is commonly done by traditional means, which tend to be time-consuming and labor-intensive. The detection relies on expert review, sometimes leading to issues of subjectivity, delayed diagnosis, and treatment. Methods: This study presents a novel Deep Learning (DL) approach, designed to improve texture-structure understanding for breast cancer classification. In the proposed approach, Res-MorphNet synergistically combines ResNet with a Morphological Texture Encoding (MTE) module to extract intricate morphologic textures, while Swin-CPSANet leverages the Swin Transformer and a Cross-Patch Spatial Aggregation (CPSA) block to enhance global context understanding and feature interaction. To the best of our knowledge, this is the first attempt to integrate morphological texture encoding with cross-patch attention within a unified CNN-Transformer fusion framework for histopathological breast cancer classification. Results: Evaluation on the BACH dataset indicates that the proposed model achieves an accuracy of 95%, a precision of 95.49%, a recall of 95%, and an F1 score of 95.05%. These results outperform traditional CNN-and transformer-based baseline models, demonstrating the effectiveness of combining morphological texture encoding with cross-patch attention in a unified framework. Conclusion: The findings indicate that the proposed architecture achieved efficient and well-balanced classification performance across multiple breast tissue classes, underscoring its potential as a valuable tool to assist in clinical diagnosis. (c) 2026 AGBM. Published by Elsevier Masson SAS. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
1) Objectives: Early and, accurate detection of lung cancer from histopathological images is time-consuming and often requires additional staining, leading to delayed diagnosis. This study aims to develop an automated and optimized framework for reliable lung cancer detection from H&E-stained histopathological images using advanced segmentation, feature optimization, and classification techniques. 2) Materials and Methods: The proposed framework begins with image preprocessing using an adaptive median filter to suppliess noise while preserving structural details. A novel Hybrid Simple Linear Iterative Clustering K-Means Fuzzy C Means (SLIC-KM-FCM) based segmentation approach is employed to extract diagnostically relevant regions. Feature dimensionality reduction is performed using the following nature-inspired optimization algorithms, namely Whale Optimization Algorithm (WOA) and Harmony Search Optimization Algorithm (HSOA). The most discriminative features are selected using Monkey Search Algorithm (MSA) and T-statistics (T-Stat). These selected features are provided as input to multiple classifiers, including SVM, KNN, RF, DT, SDC, MLP, and BLDC. Classifier performance is evaluated using standard metrics, both with and without hyper-parameter tuning using Grid Search (GS) and Stochastic Gradient Descent (SGD). 3) Results: Without hyper-parameter tuning, the SVM classifier combined with WOA-based feature extraction and MSA-baked feature selection achieved an accuracy of 87.50%. The application of hyper-parameter optimization significantly improved classification performance. The highest accuracy of 93.75% was obtained using the BLDC classifier with HSOA-based feature extraction, T-Stat feature selection, and Grid Search optimization. 4) Conclusion: The proposed hybrid optimization-driven framework effectively improves lung cancer classification from histopathological images. The integration of advanced segmentation, nature-inspired feature optimization, and hyper-parameter tuning demonstrates strong potential for developing robust computer-aided diagnostic systems in digital pathology. 2026 AGBM. Published by Elsevier Masson SAS. All rights are reserved, including those for text and data mining. Al training, and similar technologies.
Background: Inertial motion capture systems like Xsens Awinda are increasingly used for assessing movement in clinical populations. However, it remains unclear how stroke-related motor impairments affect the quality of motion capture calibration, which is critical for obtaining reliable data. Research question: Does motor impairment severity in stroke patients influence calibration quality when using the Xsens Awinda system? Methods: Forty-eight individuals with a primary stroke (median age 66 years; 21 female; FAC >= 3) performed a total of 117 motion capture assessments using the Xsens Awinda system. Calibration quality was rated using Xsens' internal quality metrics. Kruskal-Wallis tests were conducted to compare functional motor abilities, assessed using the BOOMER test and 10-meter walk time, across calibration quality levels. Results: Of the 109 valid trials, 55% achieved good, 29% acceptable, and 16% poor calibration quality. Neither the BOOMER scores nor the 10-meter walk time was significantly associated with calibration quality levels. Significance: These findings suggest that the Xsens Awinda system can be used across a range of motor impairment levels without introducing bias due to changes in calibration quality. Nonetheless, calibration may be challenging for severely affected individuals, a limitation that warrants further practical refinement of current procedures. (c) 2026 AGBM. Published by Elsevier Masson SAS. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Obstructive Sleep Apnea (OSA) is the most commonly diagnosed sleep-related breathing disorder, affecting over one billion individuals globally. While polysomnography (PSG) remains the gold standard for diagnosis, it is resource-intensive and limited in accessibility. Consequently, the development of automated diagnostic systems has emerged as a vital research area, particularly those leveraging machine learning (ML) and deep learning (DL) techniques. Objectives This narrative review aims to provide a comprehensive comparison of ML and DL-based methods for computer-assisted diagnosis of OSA in adults. The review emphasizes model architectures, performance metrics, application scenarios, and real-world deployment challenges. Special attention is given to advanced DL architectures—such as hybrid models and Transformers—as well as the potential role of wearable technologies in scalable diagnosis. Material and methods The literature search was conducted using Web of Science, IEEE Xplore, and PubMed to identify peer-reviewed articles published between 2008 and 2024. Search terms included Obstructive Sleep Apnea (OSA), k-Nearest Neighbor (k-NN), Support Vector Machine (SVM), Linear Discriminant Analysis (LDA), Logistic Regression (LR), Ensemble Algorithm, Artificial Neural Network (ANN), Convolution Neural Network (CNN), Recurrent Neural Network (RNN), Deep Belief Network (DBN), hybrid neural network, and Transformers. Results DL-based models have demonstrated superior performance over conventional ML approaches, particularly in their ability to perform automated, hierarchical feature extraction and model complex physiological patterns. Hybrid and Transformer-based networks stand out for their diagnostic accuracy and scalability. However, most models remain limited to benchmark dataset validation and lack hardware-level implementation. Key challenges include data heterogeneity, poor model interpretability, and limited clinical generalizability. Conclusion DL-driven diagnostic frameworks—especially those incorporating multimodal signals and wearable data—represent the most promising direction for achieving accurate, scalable, and accessible OSA detection. Future research should prioritize clinical validation across diverse populations, integration of explainable AI techniques, and real-world deployment through user-centered design and IoT-based wearable platforms.
The meniscus plays a vital role in knee biomechanics, contributing to shock absorption, joint stability, proprioception, and lubrication. Anterior cruciate ligament reconstruction (ACLR) aims to restore knee stability after anterior cruciate ligament (ACL) injury; however, 30-60% of patients experience concurrent or subsequent meniscal damage. Despite this, the influence of ACLR surgical parameters on meniscal biomechanics remains largely unexplored. This study investigates how four key ACLR surgical parameters-graft type, graft size, tunnel location, and pre-tensioning-affect meniscal contact forces and stress using a coupled neuromusculoskeletal-finite element (NMSKFE) modeling approach during simulated walking. NMSK-FE simulations were conducted in six participants to assess changes in meniscal-tibial contact forces and stress distributions under various ACLR configurations. While most surgical conditions restored meniscal mechanics to near-intact levels (normalized root mean square error (nRMSE) < 10%), substantial deviations were observed in certain cases, particularly in anteroposterior and mediolateral contact forces and maximum principal stress (nRMSE > 10%). Notably, posterior graft placement with zero pre-tensioning increased medial meniscus stress, potentially elevating the risk of degeneration or injury. These findings highlight the individualized nature of ACLR outcomes, influenced not only by surgical parameters but also by patient-specific factors such as knee morphology and gait patterns. This study underscores the need for pre-surgical assessments that integrate patientspecific biomechanics to optimize ACLR strategies, enhance meniscal preservation, and improve long-term knee health. By incorporating meniscal mechanics and dynamic gait analysis, this research advances personalized ACLR approaches, addressing a critical gap in the field. (c) 2026 AGBM. Published by Elsevier Masson SAS. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
BACKGROUNDS: Vascular calcification (VC) is an actively regulated dynamic process characterizing by abnormal deposition of calcium phosphate mineral in the extracellular matrix and in cells of the arterial wall. Significant advances have been made in comprehending the ferroptosis linked to VC, yet the precise molecular mechanism is still not fully understood. Interpretability and explainability of machine learning models are crucial for incorporating them into decision-making processes. We used the Shapley additive explanation (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) methods in this study to interpret and explain a random forest model in order to discover the significant attributes. METHODS: This paper employed the GEO tools to get a VC dataset. The DEGs were discovered using the EdgeR package in R to identify potential ferroptosis-associated hub genes that could be used for VC diagnosis. We used qRT-PCR and western blotting techniques to confirm the DEGs associated with ferroptosis that were discovered in the microarray data. Finally, we suggest two innovative strategies, using SHAP and LIME, to enhance interpretation. We evaluated the explanatory outcomes of the SHAP scheme with other approaches using GEO datasets. RESULTS: We uncovered 49 ferroptosis DEGs in VC, including 31 upregulated and 18 downregulated genes. The outputs obtained from the GSEA and the study of the KEGG using WebGestalt revealed that the differentially expressed genes (DEGs) related to ferroptosis are found to be involved in six paths, one of which was the Ferroptosis signaling pathway. SHAP and LIME interpretation aligned well with the interpretations provided by the current methodologies. We demonstrated the significance of TP63 and GPX2 as crucial predictive factors for VC using of suggested methodologies. Lastly, we examined the three genes identified by two machine learning models in vitro and observed that the mRNA and protein profile levels of FTH1 exhibited an elevated level and the levels of SLC3A2 and SLC7A11 exhibited a reduced level in the j3-GP-treated class in comparison to the normal class. The nomogram and 5 potential hub genes exhibited excellent predictive performance, with AUC values ranging from 0.724 to 0.969. CONCLUSIONS: Our investigation found three ferroptosis-associated potential hub genes by comprehensive exploration (FTH1, SLC3A2, and SLC7A11). In addition, we created a nomogram for VC diagnosis utilising bioinformatics and machine learning approaches (SHAP and LIME). Our methods are effective for analyzing machine learning models and may reveal the fundamental connections among variables and outputs. (c) 2025 AGBM. Published by Elsevier Masson SAS. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
Context: The imminent performance of multimodal and heterogeneous modalities to forecast the advance of Alzheimer's disease (AD) precisely is one of the key problems. Current models are usually not interpretable, time-consistent, and multimodal, making them less useful in clinical forecasting. Objective: The objective of the study is to develop a hybrid generative approach to simulate the individualized AD progression process, which can generate future anatomical and clinical states, model latent over-time dynamics, and measure the uncertainty. Methods: The proposed study suggests using a multimodal paradigm that enables a combination of Conditional Latent Diffusion Models (cLDM) and Neural Ordinary Differential Equations (ODEs). The model permits the generation of plausible future MRI, cognitive scoring, and biomarker trajectories for a patient at baseline. The ADNI dataset was evaluated with structural similarity (SSIM), clinical prediction error, and classification accuracy. Key Findings: The model provided an SSIM equal to 0.86 on synthesizing future MRI, and the MAE of MMSE prediction was equal to 1.5. It exceeded baselines in all the imaging, cognitive, and biomarker settings. The conversion of AD resulted in an accuracy of the classification of 88% with stable multimodal generalization at calibrated output of probability. Conclusion: The proposed model offers a feasible and explainable approach to the forecast of an AD trajectory, allowing realistic simulations of a digital twin and projecting its progress within a multiyear perspective. It also supports early detection, custom intervention, and uncertainty-conscious clinical decision-making. (c) 2025 AGBM. Published by Elsevier Masson SAS. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
