Gait Posture. 2025 Jul 2;122:50-57. doi: 10.1016/j.gaitpost.2025.07.005. Online ahead of print.

ABSTRACT

BACKGROUND: Measuring bilateral ground reaction forces (GRFs) and centre of pressure (COP) is essential in gait analysis, requiring subjects to step each foot sequentially onto a separate forceplate. However, this requirement often causes multiple trial attempts, especially in patients with neuromusculoskeletal disorders. Consciously targeting the forceplates could also alter walking mechanics, leading to unnatural gait patterns.

RESEARCH QUESTION: This study aimed to (1) develop a novel physics-informed residual recurrent neural network (PI-ResRNN) to predict bilateral GRF and COP during gait using data from a single forceplate and (2) evaluate its accuracy against ground truth obtained across subject groups of different ages and pathologies.

METHODS: Forceplate data from 315 participants, namely healthy participants and patients with six types of neuromusculoskeletal disorders, was collected. Data from 6765 trials was used to train and validate the PI-ResRNN model to decompose GRF and COP for each foot during the double-contact phase of walking. Model-predicted COP and GRFs were evaluated against the ground truth using root-mean-square errors (RMSE) and relative RMSE (rRMSE), respectively.

RESULTS: All predicted variables from the PI-ResRNN model demonstrated high consistency with the ground truth, with mean rRMSE values below 0.34 %, 0.38 %, and 0.56 % in the vertical, anteroposterior, and mediolateral GRFs, respectively, and mean RMSE values for COP below 3.0 mm. The model effectively identified statistical between-group differences compared with the ground truth.

SIGNIFICANCE: The proposed model provides a practical and accurate approach for obtaining bilateral GRF and COP using a single forceplate, benefiting gait analysis in populations with mobility impairments.

PMID:40618708 | DOI:10.1016/j.gaitpost.2025.07.005

Comput Biol Med. 2025 Jul 5;196(Pt A):110658. doi: 10.1016/j.compbiomed.2025.110658. Online ahead of print.

ABSTRACT

BACKGROUND: Automatic medical image segmentation techniques are vital for assisting clinicians in making accurate diagnoses and treatment plans. Although the U-shaped network (U-Net) has been widely adopted in medical image analysis, it still faces challenges in capturing long-range dependencies, particularly in complex and textured medical images where anatomical structures often blend into the surrounding background.

METHOD: To address these limitations, a novel network architecture, called recursive transformer-based U-Net (ReT-UNet), which integrates recursive feature learning and transformer technology, is proposed. One of the key innovations of ReT-UNet is the multi-scale global feature fusion (Multi-GF) module, inspired by transformer models and multi-scale pooling mechanisms. This module captures long-range dependencies, enhancing the abstraction and contextual understanding of multi-level features. Additionally, a recursive feature accumulation block is introduced to iteratively update features across layers, improving the network's ability to model spatial correlations and represent deep features in medical images. To improve sensitivity to local details, a lightweight atrous spatial pyramid pooling (ASPP) module is appended after the Multi-GF module. Furthermore, the segmentation head is redesigned to emphasize feature aggregation and fusion. During the encoding phase, a hybrid pooling layer is employed to ensure comprehensive feature sampling, thereby enabling a broader range of feature representation and improving detailed information learning.

RESULTS: Results: The proposed method has been evaluated through ablation experiments, demonstrating generally consistent performance across multiple trials. When applied to cardiac, pulmonary nodule, and polyp segmentation datasets, the method showed a reduction in mis-segmented regions. The experimental results suggest that the approach can improve segmentation accuracy and stability compared to competing state-of-the-art methods.

CONCLUSIONS: Experimental findings highlight the superiority of the proposed ReT-UNet over related methods and demonstrate its potential for applications in medical image segmentation.

PMID:40618700 | DOI:10.1016/j.compbiomed.2025.110658

Spectrochim Acta A Mol Biomol Spectrosc. 2025 Jun 27;344(Pt 1):126577. doi: 10.1016/j.saa.2025.126577. Online ahead of print.

ABSTRACT

Raman spectroscopy is a valuable analytical technique for molecular characterization, but its practical application is often restricted by low signal-to-noise ratio (SNR), especially at short integration times. These limitations are critical in time-sensitive applications where rapid spectral acquisition is required. Deep learning, in particular, U-Net-based models are becoming routine tools for denoising the spectra. However, such models are typically applied as black boxes that are only evaluated using performance metrics. In this study, we investigate how training strategies using spectra acquired at different integration times and thus varying noise levels, affect model generalization. Specifically, we compare two models trained with the same U-Net architecture: a Single-Condition (SC) model trained on spectra acquired with a single integration time, and a Multi-Condition (MC) model trained on a dataset combining multiple integration times. Besides the quantitative evaluation of the models using Root Mean Squared Error (RMSE) and Pearson Correlation Coefficient (PCC), we apply interpretability techniques to gain deeper insight into how the models process spectral data. Saliency maps and Jacobian matrix analysis enable us to visualize which spectral regions the models focus on during denoising. This study highlights that training with diverse integration times significantly improves model generalization and denoising robustness. Moreover, our use of interpretability techniques reveals how different training strategies influence model focus and decision-making, offering a novel perspective on designing explainable and noise-resilient deep learning models for Raman spectroscopy.

PMID:40618630 | DOI:10.1016/j.saa.2025.126577

J Orthop Res. 2025 Jul 6. doi: 10.1002/jor.70020. Online ahead of print.

ABSTRACT

Developmental dysplasia of the hip (DDH) includes a range of conditions caused by inadequate hip joint development. Early diagnosis is essential to prevent long-term complications. Ultrasound, particularly the Graf method, is commonly used for DDH screening, but its interpretation is highly operator-dependent and lacks standardization, especially in identifying the correct standard plane. This variability often leads to misdiagnosis, particularly among less experienced users. This study presents AI-SPS, an AI-based instant standard plane detection software for real-time hip ultrasound analysis. Using 2,737 annotated frames, including 1,737 standard and 1,000 non-standard examples extracted from 45 clinical ultrasound videos, we trained and evaluated two object detection models: SSD-MobileNet V2 and YOLOv11n. The software was further validated on an independent set of 934 additional frames (347 standard and 587 non-standard) from the same video sources. YOLOv11n achieved an accuracy of 86.3%, precision of 0.78, recall of 0.88, and F1-score of 0.83, outperforming SSD-MobileNet V2, which reached an accuracy of 75.2%. These results indicate that AI-SPS can detect the standard plane with expert-level performance and improve consistency in DDH screening. By reducing operator variability, the software supports more reliable ultrasound assessments. Integration with live systems and Graf typing may enable a fully automated DDH diagnostic workflow. Level of Evidence: Level III, diagnostic study.

PMID:40619593 | DOI:10.1002/jor.70020

Public Health Nurs. 2025 Jul 6. doi: 10.1111/phn.70003. Online ahead of print.

ABSTRACT

Community health nurses can enhance the elderly's quality of life (QoL) through personalized care, lifestyle counselling, and preventive measures. The primary objective of this study was to develop artificial intelligence (AI)-based prediction models to identify the key influencing factors that can impact the QoL in the elderly population. The estimated sample size was 500, and participants were selected using a systematic sampling technique. The pre-processing stage was applied to the primary dataset. Following this, basic machine learning (ML), deep learning (DL), and ensemble models were implemented to predict QoL. The SMOTE method was applied to balance the dataset. AdaBoost was the best-performing model, achieving an accuracy of 93.7%, with excellent recall (96.8%) and specificity (96.8%). Physical activity (48.9%) and daily activity ability (30.8%) were key QoL predictors, while regression analysis revealed physical activity (coefficient: 1.2260, p < 0.001) as a positive contributor. AI approaches help the community health nurses to predict the factors required for improving QoL early on, enabling them to provide the elderly population with the appropriate advice and future plans to manage aging challenges.

PMID:40619584 | DOI:10.1111/phn.70003

Sci Rep. 2025 Jul 6;15(1):24098. doi: 10.1038/s41598-025-10561-6.

ABSTRACT

With the exponential growth in the quantity and complexity of malware, traditional detection methods face severe challenges. This paper proposes GCSA-ResNet, a novel deep learning model that significantly enhances malware detection performance by integrating the Global Channel-Spatial Attention (GCSA) module with ResNet-50. The core innovation lies in the GCSA module, which for the first time collaboratively designs channel attention, channel shuffling, and spatial attention mechanisms to simultaneously capture local texture features and global dependency relationships in visualized malware images. Compared with existing attention models such as SE and CBAM, GCSA strengthens cross-channel information interaction through channel shuffling operations and employs spatial attention with a 7 × 7 convolutional kernel to more effectively model long-range spatial correlations. Experiments on the Malimg and Microsoft BIG 2015 datasets demonstrate that GCSA-ResNet achieves over 98.50% accuracy, representing a performance improvement of more than 0.5% compared to baseline models. Quantitative results show that the model maintains stable performance in precision, recall, and F1-score, while reducing false positive rates by 40-50%. These advancements effectively address the limitations of existing methods in feature degradation and cross-family misclassification.

PMID:40619499 | DOI:10.1038/s41598-025-10561-6

Ann Surg Oncol. 2025 Jul 6. doi: 10.1245/s10434-025-17748-1. Online ahead of print.

ABSTRACT

BACKGROUND: Early recurrence (ER) following curative-intent surgery remains a major obstacle to improving long-term outcomes in patients with pancreatic cancer (PC). The accurate preoperative prediction of ER could significantly aid clinical decision-making and guide postoperative management.

PATIENTS AND METHODS: A retrospective cohort of 493 patients with histologically confirmed PC who underwent resection was analyzed. Contrast-enhanced computed tomography (CT) images were used for tumor segmentation, followed by radiomics and deep learning feature extraction. In total, four distinct feature selection algorithms were employed. Predictive models were constructed using random forest (RF) and support vector machine (SVM) classifiers. The model performance was evaluated by the area under the receiver operating characteristic curve (AUC). A comprehensive nomogram integrating feature scores and clinical factors was developed and validated.

RESULTS: Among all of the constructed models, the Inte-SVM demonstrated superior classification performance. The nomogram, incorporating the Inte-feature score, CT-assessed lymph node status, and carbohydrate antigen 19-9 (CA19-9), yielded excellent predictive accuracy in the validation cohort (AUC = 0.920). Calibration curves showed strong agreement between predicted and observed outcomes, and decision curve analysis confirmed the clinical utility of the nomogram.

CONCLUSIONS: A CT-based deep learning radiomics nomogram enabled the accurate preoperative prediction of early recurrence in patients with pancreatic cancer. This model may serve as a valuable tool to assist clinicians in tailoring postoperative strategies and promoting personalized therapeutic approaches.

PMID:40619487 | DOI:10.1245/s10434-025-17748-1

Sci Rep. 2025 Jul 6;15(1):24104. doi: 10.1038/s41598-025-09791-5.

ABSTRACT

To tackle the challenge of discrete and complex monitoring data generated during high-speed rail tunnel construction, this study proposes a hybrid deep learning model for deformation forecasting. Using 300-hour continuous deformation records from multiple cross-sections of the G Tunnel (March 2023), a novel WOA-CNN-GRU model is developed, integrating data preprocessing, feature extraction, and prediction. The methodology incorporates quadratic exponential smoothing for outlier mitigation, followed by sequential feature extraction using convolutional neural networks (CNNs) and bidirectional gated recurrent units (GRUs). Comparative experiments demonstrate the model's superiority over conventional architectures including RNN, LSTM, GRU, and CNN-GRU. The WOA-CNN-GRU model achieves an RMSE of 0.1257 mm and a MAPE of 0.51%, significantly outperforming baseline models, with RMSE reductions of 0.9814 mm (RNN), 0.7629 mm (LSTM), 0.4188 mm (GRU), and 0.2292 mm (CNN-GRU), and MAPE improvements of 2.64%, 1.37%, 1.05%, and 0.86%, respectively. Moreover, the model exhibits robust generalization, with average absolute errors below 1 mm and relative errors under 1% across various construction methods and tunnel segments. These results provide compelling evidence for the effectiveness of hybrid intelligent models in capturing nonlinear, spatiotemporal deformation patterns in tunnel engineering. The WOA-CNN-GRU model offers practical guidance for real-time monitoring system development and risk mitigation in civil infrastructure projects.

PMID:40619456 | DOI:10.1038/s41598-025-09791-5

Sci Rep. 2025 Jul 7;15(1):24161. doi: 10.1038/s41598-025-07868-9.

ABSTRACT

Hybrid Grid System (HGS) installation in small and large residential area has major challenges due to domestic loads. Domestic loads are in different duty cycle such as (i) continuous duty i.e., vehicle charging, (ii) short time duty, (iii) periodic duty and (iv) intermittent duty. In this paper, proposed HGS comprises of Internet of Thing (IOT), Photovoltaic (PV) system and wind system (PWS) with Lithium-Phosphate battery paralleled with Super-capacitor, Deep learning controller with PWS is termed as IOT enabled PWS (IPWS). IPWS has zero export converters, reduces electricity demand on grid. Zero-export inverter avoids excess energy to grid and excess energy stored in super-capacitor. IPWS has crowd sensing for microclimatic conditions data acquisition system. Microclimatic Data is used for tuning zero export converters and Battery Management System (BMS) through IPWS. IPWS controller perform with different hybrid Deep learning algorithm such as (i) SCO-LSTM controller and JO-LSTM based BMS (ii) JO-LSTM controller and HBO-LSTM based BMS (iii) HBO-LSTM controller and SCO-LSTM based BMS. IPWS reduces time and space complexity in controller. Among the proposed methods, IPWS with JO-LSTM/ HBO-LSTM based BMS eliminates output power fluctuations and increases transient stability (TS) and damping ratio (DR). Comparative analysis for DC-link and super-capacitor in IPWS is presented. IPWS with JO-LSTM controller, super-capacitor suits for residence loads and provides 29% improved power factor, reduces harmonics 14%, DR of 6%, and low TS.

PMID:40619453 | DOI:10.1038/s41598-025-07868-9

Sci Rep. 2025 Jul 6;15(1):24133. doi: 10.1038/s41598-025-06493-w.

ABSTRACT

Fault detection is essential in guaranteeing the reliability, security, and productivity of contemporary technological and industrial systems. Faults that go unnoticed may result in disastrous failures as well as prohibitive downtimes in industries as varied as healthcare, manufacturing, and autonomous functioning. Conventional fault detection technologies tend to possess low accuracy rates, weak feature extraction, as well as limitations in generalizability across variegated faults. To overcome these shortcomings, this paper puts forward an Attention-GRU-Based Fault Classifier (AGFC-Net), which employs a sophisticated attention mechanism for improved feature extraction and correlation learning. Through the fusion of attention layers with Gated Recurrent Units (GRU), AGFC-Net is able to focus on key fault features, learn temporal dependencies, and provide better classification performance even under noisy conditions. Experimental results show that AGFC-Net attains a fault detection accuracy of 99.52%, better than conventional machine learning and deep learning algorithms. The suggested method presents a stronger, adaptive, and scalable solution for autonomous fault diagnosis, opening the door to intelligent and trustworthy fault detection systems in future power grids and industrial applications.

PMID:40619447 | DOI:10.1038/s41598-025-06493-w

Int J Retina Vitreous. 2025 Jul 6;11(1):75. doi: 10.1186/s40942-025-00695-0.

ABSTRACT

OBJECTIVES: To investigate the efficacy of an artificial intelligence (AI)-based fluid monitoring tool in optimizing the monitoring of neovascular age-related macular degeneration (nAMD) patients in a Brazilian cohort.

METHODS: This is a retrospective real-world study performed in a tertiary center in Brazil, including patients with nAMD. Spectral-domain optical coherence tomography (Spectralis, Heidelberg Engineering, Germany) images were processed at baseline and over 2 years of follow-up. Demographic and clinical data were collected. A deep learning algorithm (Fluid Monitor, RetInSight, Austria) was used to automatically quantify intraretinal fluid (IRF), subretinal fluid (SRF) and pigment epithelial detachment (PED). A longitudinal panel regression model and Log-Rank test were performed to assess the correlation between fluid volumes and treatment frequency, visual outcomes, macular atrophy (MA) and subretinal fibrosis (SF) development.

RESULTS: Ninety-nine eyes from 84 patients were included. Fifty-eight eyes were treatment-naïve. Higher IRF and PED in the 6 mm area were correlated with worse visual outcomes over a 2-year follow-up (p = 0.01 and p < 0.001, respectively). Higher IRF, SRF and PED were correlated with an increased risk of SF development (p < 0.001, p = 0.049 and p = 0.02 respectively). MA development showed no significant correlation with higher IRF, SRF nor PED in this analysis. Higher SRF volume correlated with a greater number of required intravitreal injections over 2-years.

CONCLUSION: This study investigates the multifaceted landscape of nAMD in a tertiary center in the Southeast Brazil using an AI-based fluid monitoring tool. Further studies that highlight the significance of using newly validated technologies across diverse populations worldwide will be of interest.

PMID:40619442 | DOI:10.1186/s40942-025-00695-0

Int J Behav Nutr Phys Act. 2025 Jul 6;22(1):92. doi: 10.1186/s12966-025-01795-8.

ABSTRACT

BACKGROUND: Increasing evidence positively links greenspace and physical activity (PA). However, most studies use measures of greenspace, such as satellite-based vegetation indices around the residence, which fail to capture ground-level views and day-to-day dynamic exposures, potentially misclassifying greenspace and limiting policy relevance.

METHODS: We analyzed data from the US-based Nurses' Health Study 3 Mobile Health Substudy (2018-2020). Participants wore Fitbits™ and provided smartphone global positioning system (GPS) for four 7-day periods throughout the year. Street-view greenspace (%trees, %grass, %other greenspace [flowers/plants/fields]) were derived from 2019 street-view imagery using deep-learning algorithms at a 100-meter resolution and linked to 10-minute GPS observations. Average steps-per-minute for were calculated for each 10-minute period following each GPS observation. Generalized Additive Mixed Models examined associations of street-view greenspace exposure with PA, adjusting for individual and area-level covariates. We considered effect modification by region, season, neighborhood walkability and socioeconomic status (SES), temperature, and precipitation.

RESULTS: Our sample included 335 participants (meanage= 39.4 years, n = 304,394 observations). Mean steps-per-minute per 10-minutes were 6.9 (SD = 14.6). An IQR increase (18.7%) in street-view trees was associated with a 0.36 steps-per-minute decrease (95%CI: -0.71, -0.01). In addition, an IQR increase (10.6%) in grass exposure was associated with a 0.59 steps-per-minute decrease (95% CI: -0.79, -0.40); however, the association was non-linear and flattened out after the 75th percentile of street-view grass. Conversely, an IQR increase (1.2%) in other greenspace was associated with a 1.99 steps-per-minute increase (95%CI: 0.01, 3.97). Associations were stronger in the spring and in higher SES neighborhoods, and among residents of the Northeast.

CONCLUSIONS: In this prospective cohort, momentary street-view exposure to trees and grass was inversely associated with PA, while exposure to other greenspace was positively associated. Future research should confirm these results in other populations and explore the mechanisms through which specific greenspace components influence PA.

PMID:40619433 | DOI:10.1186/s12966-025-01795-8

Magn Reson Med Sci. 2025 Jul 5. doi: 10.2463/mrms.tn.2024-0211. Online ahead of print.

ABSTRACT

This study presents a novel data augmentation approach to improve deep learning (DL)-based segmentation for 3D phase-contrast magnetic resonance angiography (PC-MRA) images affected by pulsation artifacts. Augmentation was achieved by simulating pulsation artifacts through the addition of periodic errors in k-space magnitude. The approach was evaluated on PC-MRA datasets from 16 volunteers, comparing DL segmentation with and without pulsation artifact augmentation to a level-set algorithm. Results demonstrate that DL methods significantly outperform the level-set approach and that pulsation artifact augmentation further improves segmentation accuracy, especially for images with lower velocity encoding. Quantitative analysis using Dice-Sørensen coefficient, Intersection over Union, and Average Symmetric Surface Distance metrics confirms the effectiveness of the proposed method. This technique shows promise for enhancing vascular segmentation in various anatomical regions affected by pulsation artifacts, potentially improving clinical applications of PC-MRA.

PMID:40619248 | DOI:10.2463/mrms.tn.2024-0211

Crit Rev Anal Chem. 2025 Jul 6:1-20. doi: 10.1080/10408347.2025.2528134. Online ahead of print.

ABSTRACT

Chromatographic data processing represents an increasing challenge in analytical science, particularly due to the complexity of samples and the large volume of data generated by chromatographic techniques coupled with mass spectrometry (MS). This paper presents a systematic review of technological innovations over the last six years in the development of computational tools for processing these data. The review follows the PRISMA protocol, with a search conducted across five databases (SciFinder, Scopus, Web of Science, Embase, and ScienceDirect), utilizing strategies based on indexed descriptors and Boolean combinations. Thirty-three studies were selected that met the criteria of originality, applicability, and innovation in analytical tools. The results reveal significant advancements in algorithms for peak detection, alignment, and deconvolution, with an emphasis on machine learning, deep learning, and multivariate resolution approaches. Tools such as DeepResolution, SeA-M2Net, SLAW, QPMASS, autoGCMSDataAnal, and AntDAS demonstrate automation, scalability, and higher accuracy in critical tasks such as noise filtering, baseline correction, and compound identification. The analysis also highlights the progress of open-source software, which promotes greater access and interoperability. Although challenges such as the need for annotated data and standardization remain, recent advancements signal a shift toward more robust, accessible, and adaptable solutions for chromatographic data processing, expanding the potential of analyses across various scientific and industrial contexts. In this review, 'peak deconvolution' refers to separating co-eluting chromatographic signals, while 'spectral deconvolution' denotes reconstructing pure MS/MS spectra from mixed fragments."

PMID:40618385 | DOI:10.1080/10408347.2025.2528134

J Robot Surg. 2025 Jul 6;19(1):354. doi: 10.1007/s11701-025-02501-3.

ABSTRACT

To develop a deep-learning model that improves the segmentation and detection of Facial Nerve in microsurgery, thereby increasing surgical precision and safety. We collected videos from 25 patients undergoing facial nerve decompression microsurgery. From these videos, we extracted and annotated 2724 images from 14 patients for training and validation (training set: validation set = 2452: 272). Data augmentation techniques were applied to the training set with a five-fold increase (12,260 images). To evaluate the accuracy of our model, we carefully selected and annotated 1674 images from 11 patients who had not been previously trained. We then introduced an Improved Unet model that integrates various attention mechanisms, a feature-rich skip connection mechanism, and a multi-dimensional convolutional block to overcome the challenges faced by traditional Unet models when dealing with blurred or small target images. Compared with the state-of-the-art method, our proposed model achieved the best performance. The FullGrad-generated heatmap certified that the model has learned the Facial Nerve features. The Improved Unet obtained an mIOU of 0.9165 with the validation set and an mIOU of 0.6543 with the test set. In various complex microsurgical environments including blood, occlusion, and blurriness, our model can detect and segment Facial Nerve precisely. The results demonstrate the proposed method is highly promising in providing real-time intraoperative guidance of the critical anatomical structures in microsurgery.

PMID:40618308 | DOI:10.1007/s11701-025-02501-3

Comput Methods Biomech Biomed Engin. 2025 Jul 6:1-13. doi: 10.1080/10255842.2025.2524478. Online ahead of print.

ABSTRACT

Despite the successful operation of convolutional neural networks (CNN) with obstructive sleep apnea hypopnea (OSAHS) classification, the interpretability of these models is poor. The limited capacity to understand models hinders the comprehension of end-users, including sleep specialists. At the same time, these models need labeled data; however, this is a time-consuming, labor-intensive, and costly process. Furthermore, the presence of lateral connections plays a crucial role in the field of visual neurobiology. However, up until now, there has been a lack of research on CNN that incorporate lateral connections. In light of this, we introduce a novel CNN architecture called the lateral connection CNN (LCCNN), which integrates the semantic arrangement of neurons to classify OSAHS. The LCCNN consists of several layers, including a convolution layer for extracting local features, a lateral connection layer for detecting salient wave features, a competition layer for updating filters in an unsupervised manner, and a pooling layer. The competition layer ensures that adjacent filters in each convolution layer have similar weight distribution, thus realizing the semantic arrangement of neurons in the LCCNN. We evaluate the performance of the proposed model using the University College Dublin database (UCD) and the Physionet Challenge database (PCD). The results show that the proposed model achieves high total accuracies of 97.3% (with a kappa coefficient of 0.9) on UCD and 95.6% (with a kappa coefficient of 0.83) on PCD. This work can serve as a foundation for future research on unsupervised deep learning models.

PMID:40618219 | DOI:10.1080/10255842.2025.2524478

Comput Biol Med. 2025 Jul 4;196(Pt A):110656. doi: 10.1016/j.compbiomed.2025.110656. Online ahead of print.

ABSTRACT

BACKGROUND: Intravenous infusion is a common medical intervention, but the need for constant monitoring of fluid levels increases the psychological burden on patients and the workload on healthcare providers. Intelligent infusion monitoring systems can address these issues by providing real-time alerts when the fluid level is low. However, existing methods struggle with accuracy and adaptability in detecting fluid levels under complex conditions.

METHODS: This study proposes an improved YOLOv7-tiny-based method for liquid level detection in medical infusion monitoring. To improve performance, three novel modules built upon the Extended Layers Aggregation Network (ELAN)-Dynamic Snake Convolution (DS)-ELAN, Deformable Convolution Network (DCN)-ELAN, and Partial Convolution (P)-ELAN-are proposed. These modules are designed to enhance detection accuracy for elongated structures, adapt to shape variations, and optimize computational efficiency for deployment on edge devices. The proposed method was trained on a dataset of 4296 annotated infusion bottle images captured under diverse lighting and environmental conditions. Performance was evaluated using metrics such as recall, mean average precision (mAP), and inference speed.

RESULTS: Experimental results demonstrate that the improved YOLOv7-tiny method achieves superior performance compared to the baseline YOLOv7-tiny, with a recall of 88.319% and mAP@[0.5:0.95] of 90.102%, while maintaining comparable computational complexity. Ablation studies confirm the independent contributions of each proposed module to the overall performance. The enhanced method also shows robust real-time capability on embedded devices.

CONCLUSION: The proposed method significantly improves the accuracy and usability of intelligent infusion monitoring systems, enabling real-time detection of fluid levels in medical infusion bottles. This approach reduces the workload on healthcare providers, minimizes patient risks, and demonstrates potential for broader applications in medical monitoring scenarios.

PMID:40617085 | DOI:10.1016/j.compbiomed.2025.110656

Comput Biol Med. 2025 Jul 4;196(Pt A):110708. doi: 10.1016/j.compbiomed.2025.110708. Online ahead of print.

ABSTRACT

BACKGROUND: Cardiac image analysis plays an important role in detecting and categorizing cardiovascular diseases (CVDs), such as coronary artery disease (CAD), heart failure, congenital heart defects, arrhythmias (irregular heartbeat), and valvular heart disease. The automated identification of these diseases represents a significant advancement in achieving early diagnosis and mitigating disease exacerbations. While extant methodologies offer advanced means for the automatic segmentation and identification of cardiac structures and pathologies, recent strides in deep learning (DL) and modern imaging modalities within cardiology have introduced new opportunities for researchers. This has underscored the importance of deep model compression and optimization techniques. This review comprehensively surveys recent deep learning applications in interpreting cardiac images, encompassing common imaging modalities.

METHOD: Following the PRISMA methodology, we reviewed recent research on advanced and frequently used DL architectures in cardiac image analysis, along with the most employed cardiac imaging datasets. Additionally, we analyse recent contributions focused on deep model compression in cardiac image processing tasks.

RESULTS: The application of DL techniques in cardiac image analysis has seen significant progress through the introduction of new techniques such as transformers, foundation models and compression techniques.

CONCLUSIONS: The paper concludes with a critical discussion addressing open challenges and proposing future research directions in this domain.

PMID:40617082 | DOI:10.1016/j.compbiomed.2025.110708

Artif Intell Med. 2025 Jul 1;168:103202. doi: 10.1016/j.artmed.2025.103202. Online ahead of print.

ABSTRACT

Imbalanced datasets have been a persistent challenge in bioinformatics, particularly in the context of drug-drug interaction (DDI) risk level datasets. Such imbalance can lead to biased models that perform poorly on underrepresented classes. To address this issue, one strategy is to construct a balanced dataset, while another involves employing more advanced features and models. In this study, we introduce a novel approach called DDintensity, which leverages pre-trained deep learning models as embedding generators combined with LSTM-attention models to address the imbalance in DDI risk level datasets. We tested embeddings from various domains, including images, graphs, and textual corpus. Among these, embeddings generated by BioGPT achieved the highest performance, with an Area Under the Curve (AUC) of 0.97 and an Area Under the Precision-Recall curve (AUPR) of 0.92. Our model was trained on the DDinter and further validated using the MecDDI dataset. Additionally, case studies on chemotherapeutic drugs, DB00398 (Sorafenib) and DB01204 (Mitoxantrone) used in oncology, were conducted to demonstrate the specificity and effectiveness of the this methods. Our approach demonstrates high scalability across DDI modalities, as well as the discovery of novel interactions. In summary, we introduce DDIntensity as a solution for imbalanced datasets in bioinformatics with pre-trained deep-learning embeddings.

PMID:40617062 | DOI:10.1016/j.artmed.2025.103202

Artif Intell Med. 2025 Jun 26;168:103191. doi: 10.1016/j.artmed.2025.103191. Online ahead of print.

ABSTRACT

Glioblastoma is a highly aggressive form of brain cancer characterized by rapid progression and poor prognosis. Despite advances in treatment, the underlying genetic mechanisms driving this aggressiveness remain poorly understood. In this study, we employed multimodal deep learning approaches to investigate glioblastoma heterogeneity using joint image/RNA-seq analysis. Our results reveal novel genes associated with glioblastoma. By leveraging a combination of whole-slide images and RNA-seq, as well as introducing novel methods to encode RNA-seq data, we identified specific genetic profiles that may explain different patterns of glioblastoma progression. These findings provide new insights into the genetic mechanisms underlying glioblastoma heterogeneity and highlight potential targets for therapeutic intervention. Code and data downloading instructions are available at: https://github.com/ma3oun/gbheterogeneity.

PMID:40617061 | DOI:10.1016/j.artmed.2025.103191