J Nucl Cardiol. 2025 Aug 5:102449. doi: 10.1016/j.nuclcard.2025.102449. Online ahead of print.

ABSTRACT

BACKGROUND: The REgistry of Flow and Perfusion Imaging for Artificial Intelligence with PET (REFINE PET) was established to collect multicenter PET and associated computed tomography (CT) images, together with clinical data and outcomes, into a comprehensive research resource. REFINE PET will enable validation and development of both standard and novel cardiac PET/CT processing methods.

METHODS: REFINE PET is a multicenter, international registry that contains both clinical and imaging data. The PET scans were processed using QPET software (Cedars-Sinai Medical Center, Los Angeles, CA), while the CT scans were processed using deep learning (DL) to detect coronary artery calcium (CAC). Patients were followed up for the occurrence of major adverse cardiovascular events (MACE), which include death, myocardial infarction, unstable angina, and late revascularization (>90 days from PET).

RESULTS: The REFINE PET registry currently contains data for 35,588 patients from 14 sites, with additional patient data and sites anticipated. Comprehensive clinical data (including demographics, medical history, and stress test results) were integrated with more than 2200 imaging variables across 42 categories. The registry is poised to address a broad range of clinical questions, supported by correlating invasive angiography (within 6 months of MPI) in 5972 patients and a total of 9252 major adverse cardiovascular events during a median follow-up of 4.2 years.

CONCLUSION: The REFINE PET registry leverages the integration of clinical, multimodality imaging, and novel quantitative and AI tools to advance the role of PET/CT MPI in diagnosis and risk stratification.

PMID:40774620 | DOI:10.1016/j.nuclcard.2025.102449

Int J Cardiol. 2025 Aug 5:133649. doi: 10.1016/j.ijcard.2025.133649. Online ahead of print.

ABSTRACT

Early identification of patients with acute myocardial infarction (AMI) at high risk of in-hospital mortality is crucial for optimizing treatment strategies. However, the urgent nature of these clinical scenarios often limits the ability to gather the comprehensive data necessary for accurate risk assessment. Missing data presents a substantial challenge to timely and effective risk evaluation METHODS: We developed MortiGen, an end-to-end model designed to address the dual challenges of missing data imputation and mortality risk prediction in patients with AMI. The model was trained using data from 9163 admissions in the eICU collaborative research database and externally tested on 12,166 admissions from the medical information Mart for intensive care and 2142 admissions from the Chongqing University central hospital. The performance of the MortiGen was compared with other ten models using multiple evaluation metrics RESULTS: MortiGen demonstrated robust performance in predicting in-hospital mortality using data available at 3 h, 12 h, 24 h, and throughout the entire hospital stay (ranging from 47.2 % to 89.5 % of the data available), with receiver operating characteristic curve ranging from 0.794 to 0.855. MortiGen outperformed other ten models. Notably, when relying solely on data available within the first 3 h of admission, MortiGen achieving performance better than some comparison models, which used data from the entire hospitalization CONCLUSIONS: MortiGen effectively predicts in-hospital mortality among patients with AMI, even in the early stages of hospital admission, overcoming challenges related to limited data availability due to time constraints and variability in laboratory testing conditions.

PMID:40774461 | DOI:10.1016/j.ijcard.2025.133649

Sci Rep. 2025 Aug 7;15(1):28860. doi: 10.1038/s41598-025-14408-y.

ABSTRACT

Complex artificial intelligence models, like deep neural networks, have shown exceptional capabilities to detect early-stage polyps and tumors in the gastrointestinal tract. These technologies are already beginning to assist gastroenterologists in the endoscopy suite. To understand how these complex models work and their limitations, model explanations can be useful. Moreover, medical doctors specialized in gastroenterology can provide valuable feedback on the model explanations. This study explores three different explainable artificial intelligence methods for explaining a deep neural network detecting gastrointestinal abnormalities. The model explanations are presented to gastroenterologists. Furthermore, the clinical applicability of the explanation methods from the healthcare personnel's perspective is discussed. Our findings indicate that the explanation methods are not meeting the requirements for clinical use, but that they can provide valuable information to researchers and model developers. Higher quality datasets and careful considerations regarding how the explanations are presented might lead to solutions that are more welcome in the clinic.

PMID:40775463 | DOI:10.1038/s41598-025-14408-y

Sci Rep. 2025 Aug 7;15(1):28961. doi: 10.1038/s41598-025-14886-0.

ABSTRACT

To enhance user interaction experience in furniture customization, this study optimizes an Internet of Things (IoT)-driven Artificial Intelligence (AI)-assisted design system. First, the study analyzes human-computer interaction theories in IoT environments. Second, a personalized furniture design model based on a Generative Adversarial Network (GAN) is constructed. This enhances the AI-assisted design system's ability to generate diverse design solutions while avoiding the limitations of traditional systems. Compared to other deep learning architectures (e.g., encoder-decoder networks), GAN excels in generating realistic and creative furniture design solutions. Finally, virtual reality (VR) technology is integrated to enable real-time interaction between users and customized furniture. The Kano model is used to evaluate the interactive features of the furniture. The results show that in the proposed interactive furniture customization system, female users prioritize comfort, convenient control functions, and safety. They also expect a smooth and intuitive interaction experience. Male users focus more on convenient control functions, visualization features, and safety, with Proportion of Attractive Quality (PA) scores of 60.80%, 56.32%, and 73.18%, respectively. Younger users significantly value visualization features and convenient control functions while also emphasizing safety. Middle-aged and elderly users prioritize operational functionality and comfort, with relatively lower demand for social and entertainment features. In terms of income levels, low-income users mainly focus on comfort, operational functionality, and safety, with PA values of 60.12%, 66.21%, and 72.35%, respectively. Middle-income users show higher demand for visualization features, with a PA value of 55.21%. High-income users emphasize safety and comfort more. The designed system effectively highlights the preferences of users across different genders, age groups, and income levels, enabling flexible design adjustments based on user characteristics. This method better meets the personalized needs of diverse users while addressing the limitations of traditional AI-assisted design systems in generating diverse solutions. It provides new insights for smart furniture design, enhancing adaptability and flexibility, and promoting technological innovation and interdisciplinary integration. This study holds significant academic value and practical application prospects.

PMID:40775453 | DOI:10.1038/s41598-025-14886-0

Sci Rep. 2025 Aug 7;15(1):28949. doi: 10.1038/s41598-025-14320-5.

ABSTRACT

With the growing demand for English language teaching, the efficient and accurate evaluation of students' writing ability has become a key focus in English education. This study introduces a Hybrid Feature-based Cross-Prompt Automated Essay Scoring (HFC-AES) model that leverages deep learning for intelligent text analysis. Building on traditional deep neural networks (DNNs), the model incorporates text structure features and attention mechanisms, while adversarial training is employed to optimize feature extraction and enhance cross-prompt adaptability. In the topic-independent stage, statistical methods and DNNs extract shared features for preliminary scoring. In the topic-specific stage, topic information is integrated into a hierarchical neural network to improve semantic understanding and topic alignment. Compared with existing Transformer-based scoring models, HFC-AES demonstrates superior robustness and semantic modeling capabilities. Experimental results show that HFC-AES achieves strong cross-prompt scoring performance, with an average Quadratic Weighted Kappa (QWK) of 0.856, outperforming mainstream models. Ablation studies further highlight the critical role of text structure features and attention mechanisms, particularly in improving argumentative writing assessment. Overall, HFC-AES offers effective technical support for automated essay grading, contributing to more reliable and efficient evaluation in English language teaching.

PMID:40775439 | DOI:10.1038/s41598-025-14320-5

Eur Arch Otorhinolaryngol. 2025 Aug 7. doi: 10.1007/s00405-025-09592-6. Online ahead of print.

ABSTRACT

PURPOSE: The aim of this article was to conduct a systematic review to evaluate the role and reliability of artificial intelligence (AI) and machine learning (ML) in the diagnosis, management, and potential treatment of pediatric upper airway obstruction (UAO).

METHODS: This PRISMA-based review searched PubMed, Scopus, and Web of Science for English-language studies on pediatric UAO (≤ 18 years) using AI/ML. Non-original works, unrelated topics, mixed-age studies, and those without AI/ML were excluded.

RESULTS: Out of 76 identified articles, 27 were included in the review. Most studies on AI and ML focused on pediatric obstructive sleep apnea (OSA), particularly diagnosis and severity classification.Convolutional Neural Networks (CNNs) were the most common approach, used in 29% of studies. The most frequent input modality was nocturnal blood oxygen saturation (SpO₂) signals (44%), followed by clinical parameters (14.8%), electrocardiography (ECG) (7.4%), and polysomnography (PSG) data (7.4%). Model performance varied based on input data and study design. Advanced methods for OSA show high accuracy: deep learning (88.8%), actigraphy/oximetry (96%), and smartphone oximeters (> 79%). The Sunrise algorithm reached 100% sensitivity for severe OSA. Limitations across current studies include heterogeneous patient populations, small sample sizes, and a predominant focus on obstructive sleep apnea (OSA), which may restrict the generalizability of the findings.

CONCLUSIONS: In pediatric sleep medicine, ML models have focused on diagnosis mainly using physiological signalsand XGBoost/Support Vector Machines (SVM) for clinical data. No studies addressed treatment or monitoring, and challenges like data diversity, validation, and feasibility remain.

PMID:40775390 | DOI:10.1007/s00405-025-09592-6

Alzheimers Res Ther. 2025 Aug 7;17(1):182. doi: 10.1186/s13195-025-01827-2.

ABSTRACT

BACKGROUND: Alzheimer's disease (AD) is the principal cause of dementia and requires the early diagnosis of people with mild cognitive impairment (MCI) who are at high risk of progressing. Early diagnosis is imperative for optimizing clinical management and selecting proper therapeutic interventions. Structural magnetic resonance imaging (MRI) markers have been widely investigated for predicting the conversion of MCI to AD, and recent advances in deep learning (DL) methods offer enhanced capabilities for identifying subtle neurodegenerative changes over time.

METHODS: We selected 228 MCI participants from the Alzheimer's Disease Neuroimaging Initiative (ADNI) who had at least three T1-weighted MRI scans within 18 months of baseline. MRI volumes underwent bias correction, segmentation, and radiomics feature extraction. A 3D residual network (ResNet3D) was trained using a pairwise ranking loss to capture single-timepoint risk scores. Longitudinal analyses were performed by extracting deep convolutional neural network (CNN) embeddings and gray matter radiomics for each scan, which were put into a time-aware long short-term memory (LSTM) model with an attention mechanism.

RESULTS: A single-timepoint ResNet3D model achieved modest performance (c-index ~ 0.70). Incorporating longitudinal MRI files or downstream survival models led to a pronounced prognostic improvement (c-index:0.80-0.90), but was not further improved by longitudinal radiomics data. Time-specific classification within two- and three-year and five-year windows after the last MRI acquisition showed high accuracy (AUC > 0.85). Several radiomics, including gray matter surface to volume and elongation, emerged as the most predictive features. Each SD change in the gray matter surface to volume change within the last visit was associated with an increased risk of developing AD (HR: 1.50; 95% CI: 1.25-1.79).

CONCLUSIONS: These findings emphasize the value of structural MRI within the advanced DL architectures for predicting MCI-to-AD conversion. The approach may enable earlier risk stratification and targeted interventions for individuals most likely to progress. limitations in sample size and computational resources warrant larger, more diverse studies to confirm these observations and explore additional improvements.

PMID:40775357 | DOI:10.1186/s13195-025-01827-2

Sci Rep. 2025 Aug 7;15(1):28878. doi: 10.1038/s41598-025-14927-8.

NO ABSTRACT

PMID:40775268 | DOI:10.1038/s41598-025-14927-8

Sci Rep. 2025 Aug 7;15(1):28974. doi: 10.1038/s41598-025-13689-7.

ABSTRACT

Grape disease image recognition is an important part of agricultural disease detection. Accurately identifying diseases allows for timely prevention and control at an early stage, which plays a crucial role in reducing yield losses. This study addresses the problems in grape leaf disease recognition under small-sample conditions, such as the difficulty in capturing multi-scale features, the minuteness of features, and the weak adaptability of traditional data augmentation methods. It proposes a solution that combines a multi-scale feature hybrid fusion architecture with data augmentation. The innovation of this study lies in the following four dimensions: (1) Utilize generative models to enhance the cross-category data balancing ability under small-sample conditions and enrich the sample information in the dataset. (2) Innovatively propose the LVT Block, a multi-scale information perception hybrid module based on the Ghost and Transformer structures. This module can effectively acquire and fuse multi-scale information and global information in the feature map. (3) Use the dense connection method to combine the LVT Block and the MARI Block to propose a new architecture, the DLVT Block. By fusing multi-scale information and global information, it improves the richness of feature information. It also uses the MARI to enhance the model's perception of disease areas and constructs an end-to-end lightweight model, DLVTNet, using the DLVT Block. Experiments show that this method achieves an average recognition rate of 98.48% on the New Plant Diseases Dataset. The number of parameters is reduced to 42.7% of that of MobileNetV4, and it maintains an accuracy of 96.12% in the tomato leaf disease test. This paper embeds pathological features into the generative adversarial process, which can effectively alleviate the problem of insufficient samples in intelligent agricultural detection. It provides a new method system with strong interpretability and excellent generalization performance for disease detection.

PMID:40775261 | DOI:10.1038/s41598-025-13689-7

NPJ Digit Med. 2025 Aug 7;8(1):508. doi: 10.1038/s41746-025-01861-2.

ABSTRACT

The current standard-of-care (SOC) practice for defining the clinical target volume (CTV) for radiation therapy (RT) in patients with glioblastoma still employs an isotropic 1-2 cm expansion of the T2-hyperintensity lesion, without considering the heterogeneous infiltrative nature of these tumors. This study aims to improve RT CTV definition in patients with glioblastoma by incorporating biologically relevant metabolic and physiologic imaging acquired before RT along with a deep learning model that can predict regions of subsequent tumor progression by either the presence of contrast-enhancement or T2-hyperintensity. The results were compared against two standard CTV definitions. Our multi-parametric deep learning model significantly outperformed the uniform 2 cm expansion of the T2-lesion CTV in terms of specificity (0.89 ± 0.05 vs 0.79 ± 0.11; p = 0.004), while also achieving comparable sensitivity (0.92 ± 0.11 vs 0.95 ± 0.08; p = 0.10), sparing more normal brain. Model performance was significantly enhanced by incorporating lesion size-weighted loss functions during training and including metabolic images as inputs.

PMID:40775041 | DOI:10.1038/s41746-025-01861-2

Sci Rep. 2025 Aug 7;15(1):28938. doi: 10.1038/s41598-025-14491-1.

ABSTRACT

During regional seismic monitoring, data is automatically analyzed in real-time to identify events and provide initial locations and magnitudes. Monitoring networks may apply automatic post-processing to small events (M < 3) to add and refine picks and improve the event before analyst review. Recently, machine learning algorithms, particularly for phase picking, have matured enough for use in regional monitoring systems. The Southern California Seismic Network has implemented the deep-learning picker PhaseNet in our event post-processing, resulting in about 2-3 times as many picks, particularly S phases, with slightly better pick accuracy than the previous STA/LTA picker (relative to analyst picks). These improvements have led to better epicenter accuracy. We have also developed an automatic post-processing pipeline (ST-Proc) for sub-network triggers, which are collections of nearby phase picks that the real-time system could not associate into an event. ST-Proc uses PhaseNet to find phase picks and the machine learning algorithm GaMMA to associate events. This pipeline is capable of correctly detecting events in 65-70% of triggers containing events with a low false event rate around 5%. Additionally, the GaMMA-determined epicenters are generally accurate (within a few kilometers of the final). Both pipelines have helped to reduce analyst workload and streamline event processing.

PMID:40775035 | DOI:10.1038/s41598-025-14491-1

Neuroimage Clin. 2025 Aug 6;48:103858. doi: 10.1016/j.nicl.2025.103858. Online ahead of print.

ABSTRACT

Accurate prediction of post-stroke language outcomes using machine learning offers the potential to enhance clinical treatment and rehabilitation for aphasic patients. This study of 758 English speaking stroke patients from the PLORAS project explores the impact of sample size on the performance of logistic regression and a deep learning (ResNet-18) model in predicting language outcomes from neuroimaging and impairment-relevant tabular data. We assessed the performance of both models on two key language tasks from the Comprehensive Aphasia Test: Spoken Picture Description and Naming, using a learning curve approach. Contrary to expectations, the simpler logistic regression model performed comparably or better than the deep learning model (with overlapping confidence intervals), with both models showing an accuracy plateau around 80% for sample sizes larger than 300 patients. Principal Component Analysis revealed that the dimensionality of the neuroimaging data could be reduced to as few as 20 (or even 2) dominant components without significant loss in accuracy, suggesting that classification may be driven by simple patterns such as lesion size. The study highlights both the potential limitations of current dataset size in achieving further accuracy gains and the need for larger datasets to capture more complex patterns, as some of our results indicate that we might not have reached an absolute classification performance ceiling. Overall, these findings provide insights into the practical use of machine learning for predicting aphasia outcomes and the potential benefits of much larger datasets in enhancing model performance.

PMID:40773787 | DOI:10.1016/j.nicl.2025.103858

Neural Netw. 2025 Aug 5;192:107928. doi: 10.1016/j.neunet.2025.107928. Online ahead of print.

ABSTRACT

We propose a novel Physics-Informed Neural Network (PINN) framework for solving Partial Differential Equations (PDEs) on manifolds represented by raw point clouds, without requiring any geometric priors such as level set functions or explicit surface parametrization. Unlike prior methods that assume the availability of normal vectors or a predefined representation of the surface, our approach reconstructs an implicit surface representation using normalizing flows, enabling accurate PDE solutions without requiring labeled training data. To the best of our knowledge, this is the first PINN framework that does not rely on predefined surface characteristics or supervision. Experimental results demonstrate that our method achieves high accuracy even with non-uniformly distributed and noisy point clouds, where traditional numerical approaches often fail. Furthermore, our approach exhibits significantly faster convergence compared to existing PINN-based methods that require explicit surface knowledge. This work highlights the potential of learning-based geometric methods in automating PDE simulations on arbitrary 3D surfaces.

PMID:40773782 | DOI:10.1016/j.neunet.2025.107928

Neural Netw. 2025 Aug 5;192:107903. doi: 10.1016/j.neunet.2025.107903. Online ahead of print.

ABSTRACT

Multivariate time series anomaly detection is crucial for various real-world applications to guarantee system reliability and safety. Many deep learning methods have been proposed for time series anomaly detection purposes. However, most existing approaches treat spatial and temporal features separately, limiting their capabilities to cope with complex interactions in real-world applications. In this study, we introduce MultiverseAD, a spatial-temporal synchronous attention network that incorporates causal knowledge to enhance anomaly detection performance in multivariate time series. MultiverseAD combines a dynamic spatial-temporal synchronous attention network with a static spatial-temporal causal graph. The network employs sliding graph attention to capture local and long-term dependencies, while the causal graph encodes static causal relationships to enhance the learning process. Experiments on eight public datasets have shown that MultiverseAD consistently surpasses twelve state-of-the-art models. Ablation experiment results further validate the effectiveness of both the causal graph and the synchronous attention network.

PMID:40773781 | DOI:10.1016/j.neunet.2025.107903

JMIR AI. 2025 Aug 7;4:e74053. doi: 10.2196/74053.

ABSTRACT

BACKGROUND: Overcrowded emergency rooms might degrade the quality of care and overload the clinic staff. Assessing unscheduled return visits (URVs) to the emergency department (ED) is a quality assurance procedure to identify ED-discharged patients with a high likelihood of bounce-back, to ensure patient safety, and ultimately to reduce medical costs by decreasing the frequency of URVs. The field of machine learning (ML) has evolved considerably in the past decades, and many ML applications have been deployed in various contexts.

OBJECTIVE: This study aims to develop an ML-assisted framework that identifies high-risk patients who may revisit the ED within 72 hours after the initial visit. Furthermore, this study evaluates different ML models, feature sets, and feature encoding methods in order to build an effective prediction model.

METHODS: This study proposes an ML-assisted system that extracts the features from both structured and unstructured medical data to predict patients who are likely to revisit the ED, where the structured data includes patients' electronic health records, and the unstructured data is their medical notes (subjective, objective, assessment, and plan). A 5-year dataset consisting of 184,687 ED visits, along with 324,111 historical electronic health records and the associated medical notes, was obtained from Kaohsiung Veterans General Hospital, a tertiary medical center in Taiwan, to evaluate the proposed system.

RESULTS: The evaluation results indicate that incorporating convolutional neural network-based feature extraction from unstructured ED physician narrative notes, combined with structured vital signs and demographic data, significantly enhances predictive performance. The proposed approach achieves an area under the receiver operating characteristic curve of 0.705 and a recall of 0.718, demonstrating its effectiveness in predicting URVs. These findings highlight the potential of integrating structured and unstructured clinical data to improve predictive accuracy in this context.

CONCLUSIONS: The study demonstrates that an ML-assisted framework may be applied as a decision support tool to assist ED clinicians in identifying revisiting patients, although the model's performance may not be sufficient for clinic implementation. Given the improvement in the area under the receiver operating characteristic curve, the proposed framework should be further explored as a workable decision support tool to pinpoint ED patients with a high risk of revisit and provide them with appropriate and timely care.

PMID:40773678 | DOI:10.2196/74053

PLoS One. 2025 Aug 7;20(8):e0328302. doi: 10.1371/journal.pone.0328302. eCollection 2025.

ABSTRACT

Solar greenhouse is a primary agricultural facility in northern China during winter, providing a certain level of security for the demand for vegetables and melons in the northern regions. However, there remains a lack of uniformity between crop requirements and the light and thermal environment within the planting area of the greenhouse, resulting in non-uniform growth and development of crops. The present study set out with the objective of investigating the impact of the light environment on the internal conditions of a solar greenhouse. To this end, experimental measurements were employed in conjunction with deep learning models. The results showed that rates of change in air temperature and light intensity were significantly higher in the vertical than the horizontal direction, especially below 1,800 metres, where significant differenced in temperature and light distribution existIn the horizontal direction, the impact of light distribution on soil temperature was significant within a range of less than 4,500 mm from the southern base of the greenhouse. By contrast, the impact was less pronounced within a range of 4,500 to 9,000 mm, In the temporal dimension, light variation significantly affected soil temperatures within 150 mm of the surface, but had no significant effect on temperatures within the 300-600 mm range. Similarly, light variation significantly affected temperatures within 200 mm of the inner wall surface, but had no significant effect on temperatures within the 400-800 mm range.Furthermore, vertical differences in light intensity significantly affected temperatures within the 800 mm height range from the indoor ground level, whereas the impact at other heights was less pronounced. The LSTM prediction model was highly accurate, and this study provided the necessary data and theoretical basis for regulating the light and temperature environments in solar greenhouse.

PMID:40773526 | DOI:10.1371/journal.pone.0328302

PLoS Comput Biol. 2025 Aug 7;21(8):e1012272. doi: 10.1371/journal.pcbi.1012272. Online ahead of print.

ABSTRACT

Epigenetic mechanisms play a crucial role in driving transcript expression and shaping the phenotypic plasticity of glioblastoma stem cells (GSCs), contributing to tumor heterogeneity and therapeutic resistance. These mechanisms dynamically regulate the expression of key oncogenic and stemness-associated genes, enabling GSCs to adapt to environmental cues and evade targeted therapies. Importantly, epigenetic reprogramming allows GSCs to transition between cellular states, including therapy-resistant mesenchymal-like phenotypes, underscoring the need for epigenetic-targeting strategies to disrupt these adaptive processes. Understanding these epigenetic drivers of gene expression provides a foundation for novel therapeutic interventions aimed at eradicating GSCs and improving glioblastoma outcomes. Using machine learning (ML), we employ cross-patient prediction of transcript expression in GSCs by combining epigenetic features from various sources, including ATAC-seq, CTCF ChIP-seq, RNAPII ChIP-seq, H3K27Ac ChIP-seq, and RNA-seq. We investigate different ML and deep learning (DL) models for this task and ultimately build our final pipeline using XGBoost. The model trained on one patient generalizes to other 11 patients with high performance. Notably, H3K27Ac alone from a single patient is sufficient to predict gene expression in all 11 patients. Furthermore, the distribution of H3K27Ac peaks across the genomes of all patients is remarkably similar. These findings suggest that GSCs share a common distributional pattern of enhancer activity characterized by H3K27Ac, which can be utilized to predict gene expression in GSCs across patients. In summary, while GSCs are known for their transcriptomic and phenotypic heterogeneity, we propose that they share a common epigenetic pattern of enhancer activation that defines their underlying transcriptomic expression pattern. This pattern can predict gene expression across patient samples, providing valuable insights into the biology of GSCs.

PMID:40773517 | DOI:10.1371/journal.pcbi.1012272

PLoS One. 2025 Aug 7;20(8):e0328969. doi: 10.1371/journal.pone.0328969. eCollection 2025.

ABSTRACT

Accurate and reliable skeletal motion tracking is essential for rehabilitation monitoring, enabling objective assessment of patient progress and facilitating telerehabilitation applications. Traditional marker-based motion capture systems, while highly accurate, are costly and impractical for home rehabilitation, whereas marker-less methods often suffer from depth estimation errors and occlusions. Recent studies have explored various computer vision and deep learning approaches for human pose estimation, yet challenges remain in ensuring robust depth accuracy and tracking under occlusion conditions. This study proposes a three-dimensional human skeleton tracking system for upper limb activities that integrates temporal and spatial synchronization to improve depth estimation accuracy for rehabilitation exercises. The proposed system combines a 90° secondary camera to compensate for the depth prediction inaccuracies inherent in single-camera systems, reducing error margins by up to 0.4 m. In addition, a linear regression-based depth error correction model is implemented to refine depth coordinates, further improving tracking precision. The Kalman filtering framework is employed to enhance temporal consistency, allowing real-time interpolation of missing joint positions. Experimental results demonstrate that the proposed method significantly reduces depth estimation errors of the elbow and wrist joint (p < 0.001) compared to single camera setups, particularly in scenarios involving occlusions and non-frontal perspectives. This study provides a cost-effective and scalable solution for remote patient monitoring and motor function evaluation.

PMID:40773500 | DOI:10.1371/journal.pone.0328969

PLoS One. 2025 Aug 7;20(8):e0329555. doi: 10.1371/journal.pone.0329555. eCollection 2025.

ABSTRACT

The rapid development of Internet of Things (IoT) technology and deep learning has propelled the deployment of vision-based fire detection algorithms on edge devices, significantly exacerbating the trade-off between accuracy and inference speed under hardware resource constraints. To address this issue, this paper proposes FCMI-YOLO, a real-time fire detection algorithm optimized for edge devices. Firstly, the FasterNext module is proposed to reduce computational cost and enhance detection precision through lightweight design. Secondly, the Cross-Scale Feature Fusion Module (CCFM) and the Mixed Local Channel Attention (MLCA) mechanism are incorporated into the neck network to improve detection performance for small fire targets and reduce resource consumption. Finally, the Inner-DIoU loss function is proposed to optimize bounding box regression. Experimental results on a custom fire dataset demonstrate that FCMI-YOLO increases mAP@50 by 1.5%, reduces parameters by 40%, and lowers GFLOPs to 28.9% of YOLOv5s, demonstrating its practical value for real-time fire detection in edge scenarios with limited computational resources. The core code and dataset are available at https://github.com/ JunJieLu20230823/code.git.

PMID:40773480 | DOI:10.1371/journal.pone.0329555

PLoS One. 2025 Aug 7;20(8):e0328899. doi: 10.1371/journal.pone.0328899. eCollection 2025.

ABSTRACT

Logistics networks are becoming increasingly complex and rely more heavily on real-time vehicle data, necessitating intelligent systems to monitor driver behavior and identify route anomalies. Traditional techniques struggle to capture the dynamic spatiotemporal relationships that define driver actions, route deviations, and operational inefficiencies in big fleets. This paper introduces GNN-RMNet, a hybrid deep learning system that combines GNN, ResNet, and MobileNet for interpretable, scalable, and efficient driver behavior profiling and route anomaly detection. GNN-RMNet utilizes spatiotemporal GPS trajectories and vehicle sensor streams to learn contextual and relational patterns from structured driving data in real time, thereby identifying dangerous driving and route violations. On a real-world GPS-vehicle sensor dataset, the proposed model achieves 98% accuracy, 97% recall, an F1-score of 97.5%, and domain-specific measures like Anomaly Detection Precision (96%) and Route Deviation Sensitivity (95%). Modular design offloads ResNet-GNN analytics to edge nodes while preserving MobileNet components for on-vehicle inference, resulting in reduced inference latency (32 ms). Comparing GNN-RMNet against baseline, ensemble, and hybrid models shows its accuracy, efficiency, and generalization advantages. Computational feasibility, anomaly scoring interpretability, and future deployment concerns, including cybersecurity, data privacy, and multimodal sensor integration, are all covered. For real-time fleet safety management and secure, intelligent, and context-aware logistics, GNN-RMNet seems promising. The framework incorporates multimodal, privacy-aware, and scalable driver analytics, enabling its use in intelligent transportation systems and urban logistics infrastructures.

PMID:40773479 | DOI:10.1371/journal.pone.0328899