Sci Rep. 2025 Feb 26;15(1):6911. doi: 10.1038/s41598-025-89709-3.

ABSTRACT

Wearable devices face a significant challenge in balancing battery life with performance, often leading to frequent recharging and reduced user satisfaction. In this paper, we introduce the SmartAPM (Smart Adaptive Power Management) framework, a novel approach that leverages deep reinforcement learning (DRL) to optimize power management in wearable devices. The key objective of SmartAPM is to prolong battery life while enhancing user experience through dynamic adjustments to specific usage patterns. We compiled a comprehensive dataset by integrating user activity data, sensor readings, and power consumption metrics from various sources, including WISDM, UCI HAR, and ExtraSensory. Synthetic power profiles and device specifications were incorporated into the dataset to enhance training. SmartAPM employs a multi-agent deep reinforcement learning framework that combines on-device and cloud-based learning techniques, as well as transfer learning, to enhance personalization. Simulations on wearable devices demonstrate that SmartAPM can extend battery life by 36% compared to traditional methods, while also increasing user satisfaction by 25%. The system adapts to new usage patterns within 24 h and utilizes less than 5% of the device's resources. SmartAPM has the potential to revolutionize energy management in wearable devices, inspiring a new era of battery efficiency and user satisfaction.

PMID:40011572 | DOI:10.1038/s41598-025-89709-3

Sci Rep. 2025 Feb 26;15(1):6916. doi: 10.1038/s41598-025-91293-5.

ABSTRACT

As the quantity of motor vehicles and drivers experiences a continuous upsurge, the road driving environment has grown progressively more complex. This complexity has led to a concomitant increase in the probability of traffic accidents. Ample research has demonstrated that distracted driving constitutes a primary human - related factor precipitating these accidents. Therefore, the real - time monitoring and issuance of warnings regarding distracted driving behaviors are of paramount significance. In this research, an intelligent driver state monitoring methodology founded on the RES - SE - CNN model architecture is proposed. When compared with three classical models, namely VGG19, DenseNet121, and ResNet50, the experimental outcomes indicate that the RES - SE - CNN model exhibits remarkable performance in the detection of driver distraction. Specifically, it attains a correct recognition rate of 97.28%. The RES - SE - CNN network architecture model is characterized by lower memory occupancy, rendering it more amenable to deployment on vehicle mobile terminals. This study validates the potential application of the intelligent driver distraction monitoring model, which is based on transfer learning, within the actual driving environment.

PMID:40011564 | DOI:10.1038/s41598-025-91293-5

J Mol Graph Model. 2025 Feb 21;137:108985. doi: 10.1016/j.jmgm.2025.108985. Online ahead of print.

ABSTRACT

Recent advancements in Artificial Intelligence (AI) and deep learning have had a significant impact on drug discovery. The prediction of molecular properties, such as toxicity and blood-brain barrier (BBB) permeability, is crucial for accelerating drug development. The accuracy of these predictions largely depends on the selection of molecular descriptors. Self-supervised learning (SSL) has gained prominence due to its strong generalization capabilities. Graph contrastive learning (GCL), a type of SSL, is particularly useful in this context. Current GCL methods for molecular graphs use various data augmentation techniques, which may potentially alter the inherent structure of molecules. Additionally, traditional single-perspective representations do not fully capture the complexity of molecules. We present RFA-FFM (Fragment-level Feature Fusion Method using Retrosynthetic Fragmentation Algorithm), which integrates molecular representations from multiple perspectives. This method employs two strategies: (1) contrasting chemical information from fragments generated by two retrosynthetic methods to provide detailed contrastive insights; (2) fusing chemical information at different levels of molecular hierarchy, including the entire molecule and its fragments. Experiments show that RFA-FFM enhances the performance of deep learning models in predicting molecular properties, improving ROC-AUC scores by 0.3 %-2.6 % compared to baselines across four classification benchmarks. Case studies on hepatitis B virus datasets demonstrate that RFA-FFM outperforms baselines by 7 %-11 %. When compared to BPE and CC-Single fragmentation algorithms, RFA-FFM shows a 2 %-4 % improvement in BBB permeability tasks, thus demonstrating its effectiveness in predicting molecular properties.

PMID:40009893 | DOI:10.1016/j.jmgm.2025.108985

J Neural Eng. 2025 Feb 26. doi: 10.1088/1741-2552/adba8c. Online ahead of print.

ABSTRACT

OBJECTIVE: Functional brain network (FBN) methods are commonly integrated with deep learning (DL) models for EEG analysis. Typically, an FBN is constructed to extract features from EEG data, which are then fed into a DL model for further analysis. Beyond this two-step approach, there is potential to embed FBN construction directly within DL models as a feature extraction module, enabling the models to learn EEG representations end-to-end while incorporating insights from FBNs. However, a critical prerequisite is whether DL models can effectively learn the FBN construction process.

APPROACH: To address this, we propose using DL models to learn FBN matrices derived from EEG data. The ability of DL models to accurately reproduce these matrices would validate their capacity to learn the FBN construction process. This approach is tested on two publicly available EEG datasets, utilizing seven DL models to learn four representative FBN matrices. Model performance is assessed through mean squared error (MSE), Pearson correlation coefficient (Corr), and concordance correlation coefficient (CCC) between predicted and actual matrices.

MAIN RESULTS: The results show that DL models achieve low MSE and relatively high Corr and CCC values when learning the Coherence network. Visualizations of predicted and error matrices reveal that while DL models capture the general structure of all four FBNs, certain regions remain difficult to model accurately. Additionally, a paired t-test comparing global efficiency and nodal degree between predicted and actual networks indicates that most predicted networks significantly differ from the actual networks (p < 0.05).

SIGNIFICANCE: These findings suggest that while DL&#xD;models can learn the connectivity relationships of certain FBNs, they struggle to capture the intrinsic topological structures. This highlights the irreplaceability of traditional FBN methods in EEG analysis and underscores the need for hybrid strategies that combine FBN methods with DL models for a more comprehensive analysis.

PMID:40009886 | DOI:10.1088/1741-2552/adba8c

J Neural Eng. 2025 Feb 26. doi: 10.1088/1741-2552/adba8a. Online ahead of print.

ABSTRACT

Brain-computer interface (BCI) is a cutting-edge technology that enables interaction with external devices by decoding human intentions, and is highly valuable in the fields of medical rehabilitation and human-robot collaboration. The technique of decoding motor intent for motor execution (ME) based on electroencephalographic (EEG) signals is in the feasibility study stage. There are still insufficient studies on the accuracy of motor execution EEG signal recognition in between-subjects classification to reach the level of realistic applications. This paper aims to investigate EEG signal-driven hand movement recognition by analyzing low-frequency time-domain (LFTD) information. Experiments with four types of hand movements, two force parameter (extraction and pushing) tasks, and a four-target directional displacement task were designed and executed, and the EEG data from thirteen healthy volunteers was collected. Sliding window approach is used to expand the dataset in order to address the issue of EEG signal overfitting. Furtherly, CNN-BiLSTM model, an end-to-end serial combination of a Bidirectional Long Short-Term Memory Network (BiLSTM) and Convolutional Neural Network (CNN) is constructed to classify the raw EEG data to recognize the hand movement. According to experimental data, the model is able to categorize four types of hand movements, extraction movements, pushing movements, and four target direction displacement movements with an accuracy of 99.14%±0.49%, 99.29%±0.11%, 99.23%±0.60%, and 98.11%± 0.23%, respectively. Furthermore, comparative tests conducted with alternative deep learning models (LSTM, CNN, EEGNet, CNN-LSTM) demonstrates that the CNN-BiLSTM model is with practicable accuracy in terms of EEG-based hand movement recognition and its parameter decoding.

PMID:40009879 | DOI:10.1088/1741-2552/adba8a

Invest Radiol. 2025 Feb 25. doi: 10.1097/RLI.0000000000001168. Online ahead of print.

ABSTRACT

OBJECTIVE: The 3-dimensional (3D) double echo steady state (DESS) magnetic resonance imaging sequence can image knee cartilage with high, isotropic resolution, particularly at high and ultra-high field strengths. Advanced undersampling techniques with high acceleration factors can provide the short acquisition times required for clinical use. However, the optimal undersampling scheme and its limits are unknown.

MATERIALS AND METHODS: High-resolution isotropic (reconstructed voxel size: 0.3 × 0.3 × 0.3 mm3) 3D DESS images of 40 knees in 20 volunteers were acquired at 7 T with varying undersampling factors (R = 4-30) and schemes (regular: GRAPPA, CAIPIRINHA; incoherent: compressed sensing [CS]), whereas the remaining imaging parameters were kept constant. All imaging data were reconstructed with deep learning (DL) algorithms. Three readers rated image quality on a 4-point Likert scale. Four-fold accelerated GRAPPA was used as reference standard. Incidental cartilage lesions were graded on a modified Whole-Organ Magnetic Resonance Imaging Score (WORMS). Friedman's analysis of variance characterized rating differences. The interreader agreement was assessed using κ statistics.

RESULTS: The quality of 16-fold accelerated CS images was not rated significantly different from that of 4-fold accelerated GRAPPA and 8-fold accelerated CAIPIRINHA images, whereas the corresponding data were acquired 4.5 and 2 times faster (01:12 min:s) than in 4-fold accelerated GRAPPA (5:22 min:s) and 8-fold accelerated CAIPIRINHA (2:22 min:s) acquisitions, respectively. Interreader agreement for incidental cartilage lesions was almost perfect for 4-fold accelerated GRAPPA (κ = 0.91), 8-fold accelerated CAIPIRINHA (κ = 0.86), and 8- to 16-fold accelerated CS (κ = 0.91).

CONCLUSIONS: Our results suggest significant advantages of incoherent versus regular undersampling patterns for high-resolution 3D DESS cartilage imaging with high acceleration factors. The combination of CS undersampling with DL reconstruction enables fast, isotropic, high-resolution acquisitions without apparent impairment of image quality. Since DESS specific absorption rate values tend to be moderate, CS DESS with DL reconstruction promises potential for high-resolution assessment of cartilage morphology and other musculoskeletal anatomies at 7 T.

PMID:40009727 | DOI:10.1097/RLI.0000000000001168

PLoS One. 2025 Feb 26;20(2):e0318992. doi: 10.1371/journal.pone.0318992. eCollection 2025.

ABSTRACT

In the acquisition of Magnetic Resonance (MR) images shorter scan times lead to higher image noise. Therefore, automatic image denoising using deep learning methods is of high interest. In this work, we concentrate on image denoising of MR images containing line-like structures such as roots or vessels. In particular, we investigate if the special characteristics of these datasets (connectivity, sparsity) benefit from the use of special loss functions for network training. We hereby translate the Perceptual Loss to 3D data by comparing feature maps of untrained networks in the loss function. We tested the performance of untrained Perceptual Loss (uPL) on 3D image denoising of MR images displaying brain vessels (MR angiograms - MRA) and images of plant roots in soil. In this study, 536 MR images of plant roots in soil and 450 MRA images are included. The plant root dataset is split to 380, 80, and 76 images for training, validation, and testing. The MRA dataset is split to 300, 50, and 100 images for training, validation, and testing. We investigate the impact of various uPL characteristics such as weight initialization, network depth, kernel size, and pooling operations on the results. We tested the performance of the uPL loss on four Rician noise levels (1%, 5%, 10%, and 20%) using evaluation metrics such as the Structural Similarity Index Metric (SSIM). Our results are compared with the frequently used L1 loss for different network architectures. We observe, that our uPL outperforms conventional loss functions such as the L1 loss or a loss based on the Structural Similarity Index Metric (SSIM). For MRA images the uPL leads to SSIM values of 0.93 while L1 and SSIM loss led to SSIM values of 0.81 and 0.88, respectively. The uPL network's initialization is not important (e.g. for MR root images SSIM differences of 0.01 occur across initializations, while network depth and pooling operations impact denoising performance slightly more (SSIM of 0.83 for 5 convolutional layers and kernel size 3 vs. 0.86 for 5 convolutional layers and kernel size 5 for the root dataset). We also find that small uPL networks led to better or comparable results than using large networks such as VGG (e.g. SSIM values of 0.93 and 0.90 for a small and a VGG19 uPL network in the MRA dataset). In summary, we demonstrate superior performance of our loss for both datasets, all noise levels, and three network architectures. In conclusion, for images containing line-like structures, uPL is an alternative to other loss functions for 3D image denoising. We observe that small uPL networks have better or equal performance than very large network architectures while requiring lower computational costs and should therefore be preferred.

PMID:40009630 | DOI:10.1371/journal.pone.0318992

PLoS One. 2025 Feb 26;20(2):e0310992. doi: 10.1371/journal.pone.0310992. eCollection 2025.

ABSTRACT

With the dramatic increase in the number of published papers and the continuous progress of deep learning technology, the research on name disambiguation is at a historic peak, the number of paper authors is increasing every year, and the situation of authors with the same name is intensifying, therefore, it is a great challenge to accurately assign the newly published papers to their respective authors. The current mainstream methods for author disambiguation are mainly divided into two methods: feature-based clustering and connection-based clustering, but none of the current mainstream methods can efficiently deal with the author name disambiguation problem, For this reason, this paper proposes the author name ablation method based on the relational graph heterogeneous attention neural network, first extract the semantic and relational information of the paper, use the constructed graph convolutional embedding module to train the splicing to get a better feature representation, and input the constructed network to get the vector representation. As the existing graph heterogeneous neural network can not learn different types of nodes and edge interaction, add multiple attention, design ablation experiments to verify its impact on the network. Finally improve the traditional hierarchical clustering method, combined with the graph relationship and topology, using training vectors instead of distance calculation, can automatically determine the optimal k-value, improve the accuracy and efficiency of clustering. The experimental results show that the average F1 value of this paper's method on the Aminer dataset is 0.834, which is higher than other mainstream methods.

PMID:40009590 | DOI:10.1371/journal.pone.0310992

Eur J Nucl Med Mol Imaging. 2025 Feb 26. doi: 10.1007/s00259-025-07165-7. Online ahead of print.

ABSTRACT

PURPOSE: Respiratory motion during PET acquisition may produce lesion blurring. Ultra-fast 20-second breath-hold (U2BH) PET reduces respiratory motion artifacts, but the shortened scanning time increases statistical noise and may affect diagnostic quality. This study aims to denoise the U2BH PET images using a deep learning (DL)-based method.

METHODS: The study was conducted on two datasets collected from five scanners where the first dataset included 1272 retrospectively collected full-time PET data while the second dataset contained 46 prospectively collected U2BH and the corresponding full-time PET/CT images. A robust and data-efficient DL method called mask vision transformer (Mask-ViT) was proposed which, after fine-tuned on a limited number of training data from a target scanner, was directly applied to unseen testing data from new scanners. The performance of Mask-ViT was compared with state-of-the-art DL methods including U-Net and C-Gan taking the full-time PET images as the reference. Statistical analysis on image quality metrics were carried out with Wilcoxon signed-rank test. For clinical evaluation, two readers scored image quality on a 5-point scale (5 = excellent) and provided a binary assessment for diagnostic quality evaluation.

RESULTS: The U2BH PET images denoised by Mask-ViT showed statistically significant improvement over U-Net and C-Gan on image quality metrics (p < 0.05). For clinical evaluation, Mask-ViT exhibited a lesion detection accuracy of 91.3%, 90.4% and 91.7%, when it was evaluated on three different scanners.

CONCLUSION: Mask-ViT can effectively enhance the quality of the U2BH PET images in a data-efficient generalization setup. The denoised images meet clinical diagnostic requirements of lesion detectability.

PMID:40009163 | DOI:10.1007/s00259-025-07165-7

Abdom Radiol (NY). 2025 Feb 26. doi: 10.1007/s00261-025-04849-4. Online ahead of print.

ABSTRACT

BACKGROUND: Hepatocellular carcinoma (HCC) exhibits a high recurrence rate, and early recurrence significantly jeopardizes patient prognosis, necessitating reliable methods for early recurrence prediction.

METHODS: Utilizing multi-institutional data and integrating deep learning (DL) techniques, we established a neural network based on DenseNet capable of concurrently processing patients' triphasic enhanced CT scans. By incorporating an attention mechanism, the model automatically focuses on regions that significantly impact patient survival. Performance metrics were first evaluated using the concordance index (C-index), calibration curves, and decision curves based on the training and validation cohorts. Finally, class activation map (CAM) techniques were employed to visualize the regions of interest identified by the model. After model construction, five-fold cross-validation was performed to assess overfitting risks and further evaluate model stability.

RESULTS: We retrospectively collected data from 302 cases across five centers, including patients who underwent Partial Hepatectomy between December 2016 and December 2022. During model development, 180 patients from Institution I formed the training cohort, while the remaining patients comprised the validation cohort. The area under the ROC curve (AUC) for two-year outcomes was 0.797 in the validation cohort. Calibration curves, survival curves, and decision curve analysis (DCA) demonstrated the model's robust performance. CAMs revealed that the model primarily focuses on intra-abdominal solid organs, consistent with clinical experience. After model development, datasets were merged for cross-validation. The best model achieved a C-index of 0.774 in the validation cohort, with five-fold cross-validation yielding an average C-index of 0.778. The 95% confidence interval (CI) for the C-index, derived from cross-validation, ranged from 0.762 to 0.793.

CONCLUSION: Our DL-based enhanced CT network shows promise in predicting early recurrence in patients, representing a potential new strategy for early recurrence prediction in HCC.

PMID:40009155 | DOI:10.1007/s00261-025-04849-4

Med Biol Eng Comput. 2025 Feb 26. doi: 10.1007/s11517-025-03327-9. Online ahead of print.

ABSTRACT

Implant dentistry is the standard of care for the replacement of missing teeth. It is a complex process where cone-beam computed tomography (CBCT) images are analyzed by the dentist to determine the implants' length, diameter, and position, and angulation diameter, position, and angulation taking into consideration the prosthodontic treatment plan, bone morphology, and position of adjacent vital anatomical structures. This traditional procedure is time-consuming and relies heavily on the dentist's knowledge and expertise, which makes it subject to human errors. This study presents a two-stage framework for the placement of dental implants. The first stage utilizes YOLOv11 for the detection of fiducial markers and adjacent bone within 2D slices of 3D CBCT images. In the second stage, classification and regression are applied to extract the apical and occlusal coordinates of the implants and to predict the implants' intra-osseous length and intra-osseous diameter. YOLOv11 achieved a 59% F-score in the marker detection phase. The mean absolute error for the implant position prediction ranged from 11.931 to 15.954. The classification of the intra-osseous diameter showed 76% accuracy, and the intra-osseous length showed an accuracy of 59%. Our results were reviewed by an expert prosthodontist and deemed promising.

PMID:40009142 | DOI:10.1007/s11517-025-03327-9

Radiol Artif Intell. 2025 Feb 26:e240303. doi: 10.1148/ryai.240303. Online ahead of print.

ABSTRACT

"Just Accepted" papers have undergone full peer review and have been accepted for publication in Radiology: Artificial Intelligence. This article will undergo copyediting, layout, and proof review before it is published in its final version. Please note that during production of the final copyedited article, errors may be discovered which could affect the content. Purpose To develop a deep learning (DL) model that derives aligned strain values from cine (noncontrast) cardiac MRI and evaluate performance of these values to predict myocardial fibrosis in patients with Duchenne muscular dystrophy (DMD). Materials and Methods This retrospective study included 139 male patients with DMD who underwent cardiac MRI examinations at a single center between February 2018 and April 2023. A DL pipeline was developed to detect five key frames throughout the cardiac cycle, and respective dense deformation fields, allowing for phase-specific strain analysis across patients and from one key frame to the next. Effectiveness of these strain values in identifying abnormal deformations associated with fibrotic segments was evaluated in 57 patients (15.2 ± 3.1 years), and reproducibility was assessed in 82 patients (12.8 ± 2.7 years), comparing our method with existing feature-tracking and DL-based methods. Statistical analysis compared strain values using t tests, mixed models, and 2000+ ML models, reporting accuracy, F1 score, sensitivity, and specificity. Results DL-based aligned strain identified five times more differences (29 versus 5, P < .01) between fibrotic and nonfibrotic segments compared with traditional strain values and identified abnormal diastolic deformation patterns often missed by traditional methods. Additionally, aligned strain values enhanced performance of predictive models for myocardial fibrosis detection, improving specificity by 40%, overall accuracy by 17%, and accuracy patients with preserved ejection fraction by 61%. Conclusion The proposed aligned strain technique enables motion-based detection of myocardial dysfunction on contrast free cardiac MRI, facilitating detailed interpatient strain analysis, and allowing precise tracking of disease progression in DMD. ©RSNA, 2025.

PMID:40008976 | DOI:10.1148/ryai.240303

J Comput Assist Tomogr. 2025 Feb 26. doi: 10.1097/RCT.0000000000001734. Online ahead of print.

ABSTRACT

The development of novel image reconstruction algorithms has been pivotal in enhancing image quality and reducing radiation dose in computed tomography (CT) imaging. Traditional techniques like filtered back projection perform well under ideal conditions but fail to generate high-quality images under low-dose, sparse-view, and limited-angle conditions. Iterative reconstruction methods improve upon filtered back projection by incorporating system models and assumptions about the patient, yet they can suffer from patchy image textures. The emergence of artificial intelligence (AI), particularly deep learning, has further advanced CT reconstruction. AI techniques have demonstrated great potential in reducing radiation dose while preserving image quality and noise texture. Moreover, AI has exhibited unprecedented performance in addressing challenging CT reconstruction problems, including low-dose CT, sparse-view CT, limited-angle CT, and interior tomography. This review focuses on the latest advances in AI-based CT reconstruction under these challenging conditions.

PMID:40008975 | DOI:10.1097/RCT.0000000000001734

J Chem Inf Model. 2025 Feb 26. doi: 10.1021/acs.jcim.4c02264. Online ahead of print.

ABSTRACT

Osteoporosis is a systemic microstructural degradation of bone tissue, often accompanied by fractures, pain, and other complications, resulting in a decline in patients' life quality. In response to the increased incidence of osteoporosis, related drug discovery has attracted more and more attention, but it is often faced with challenges due to long development cycle and high cost. Deep learning with powerful data processing capabilities has shown significant advantages in the field of drug discovery. With the development of technology, it is more and more applied to all stages of drug discovery. In particular, large models, which have been developed rapidly recently, provide new methods for understanding disease mechanisms and promoting drug discovery because of their large parameters and ability to deal with complex tasks. This review introduces the traditional models and large models in the deep learning domain, systematically summarizes their applications in each stage of drug discovery, and analyzes their application prospect in osteoporosis drug discovery. Finally, the advantages and limitations of large models are discussed in depth, in order to help future drug discovery.

PMID:40008920 | DOI:10.1021/acs.jcim.4c02264

Chin Med J (Engl). 2025 Feb 26. doi: 10.1097/CM9.0000000000003489. Online ahead of print.

ABSTRACT

Artificial intelligence (AI), particularly deep learning, has demonstrated remarkable performance in medical imaging across a variety of modalities, including X-ray, computed tomography (CT), magnetic resonance imaging (MRI), ultrasound, positron emission tomography (PET), and pathological imaging. However, most existing state-of-the-art AI techniques are task-specific and focus on a limited range of imaging modalities. Compared to these task-specific models, emerging foundation models represent a significant milestone in AI development. These models can learn generalized representations of medical images and apply them to downstream tasks through zero-shot or few-shot fine-tuning. Foundation models have the potential to address the comprehensive and multifactorial challenges encountered in clinical practice. This article reviews the clinical applications of both task-specific and foundation models, highlighting their differences, complementarities, and clinical relevance. We also examine their future research directions and potential challenges. Unlike the replacement relationship seen between deep learning and traditional machine learning, task-specific and foundation models are complementary, despite inherent differences. While foundation models primarily focus on segmentation and classification, task-specific models are integrated into nearly all medical image analyses. However, with further advancements, foundation models could be applied to other clinical scenarios. In conclusion, all indications suggest that task-specific and foundation models, especially the latter, have the potential to drive breakthroughs in medical imaging, from image processing to clinical workflows.

PMID:40008785 | DOI:10.1097/CM9.0000000000003489

Nat Mach Intell. 2025;7(2):232-243. doi: 10.1038/s42256-024-00971-y. Epub 2025 Jan 28.

ABSTRACT

Accurate in silico determination of CD8+ T cell epitopes would greatly enhance T cell-based vaccine development, but current prediction models are not reliably successful. Here, motivated by recent successes applying machine learning to complex biology, we curated a dataset of 651,237 unique human leukocyte antigen class I (HLA-I) ligands and developed MUNIS, a deep learning model that identifies peptides presented by HLA-I alleles. MUNIS shows improved performance compared with existing models in predicting peptide presentation and CD8+ T cell epitope immunodominance hierarchies. Moreover, application of MUNIS to proteins from Epstein-Barr virus led to successful identification of both established and novel HLA-I epitopes which were experimentally validated by in vitro HLA-I-peptide stability and T cell immunogenicity assays. MUNIS performs comparably to an experimental stability assay in terms of immunogenicity prediction, suggesting that deep learning can reduce experimental burden and accelerate identification of CD8+ T cell epitopes for rapid T cell vaccine development.

PMID:40008296 | PMC:PMC11847706 | DOI:10.1038/s42256-024-00971-y

Phys Imaging Radiat Oncol. 2025 Feb 1;33:100719. doi: 10.1016/j.phro.2025.100719. eCollection 2025 Jan.

ABSTRACT

Synthetic Computed Tomography (sCT) is required to provide electron density information for MR-only radiotherapy. Deep-learning (DL) methods for sCT generation show improved dose congruence over other sCT generation methods (e.g. bulk density). Using 30 female pelvis datasets to train a cycleGAN-inspired DL model, this study found mean dose differences between a deformed planning CT (dCT) and sCT were 0.2 % (D98 %). Three Dimensional Gamma analysis showed a mean of 90.4 % at 1 %/1mm. This study showed accurate sCTs (dose) can be generated from routinely available T2 spin echo sequences without the need for additional specialist sequences.

PMID:40008279 | PMC:PMC11851199 | DOI:10.1016/j.phro.2025.100719

Ther Adv Drug Saf. 2025 Feb 24;16:20420986251321704. doi: 10.1177/20420986251321704. eCollection 2025.

ABSTRACT

Artificial intelligence (AI) is transforming medication research and development, giving clinicians new treatment options. Over the past 30 years, machine learning, deep learning, and neural networks have revolutionized drug design, target identification, and clinical trial predictions. AI has boosted pharmaceutical R&D (research and development) by identifying new therapeutic targets, improving chemical designs, and predicting complicated protein structures. Furthermore, generative AI is accelerating the development and re-engineering of medicinal molecules to cater to both common and rare diseases. Although, to date, no AI-generated medicinal drug has been FDA-approved, HLX-0201 for fragile X syndrome and new molecules for idiopathic pulmonary fibrosis have entered clinical trials. However, AI models are generally considered "black boxes," making their conclusions challenging to understand and limiting the potential due to a lack of model transparency and algorithmic bias. Despite these obstacles, AI-driven drug discovery has substantially reduced development times and costs, expediting the process and financial risks of bringing new medicines to market. In the future, AI is expected to continue to impact pharmaceutical innovation positively, making life-saving drug discoveries faster, more efficient, and more widespread.

PMID:40008227 | PMC:PMC11851753 | DOI:10.1177/20420986251321704

JAMIA Open. 2025 Jan 31;8(1):ooaf007. doi: 10.1093/jamiaopen/ooaf007. eCollection 2025 Feb.

ABSTRACT

OBJECTIVES: The automatic detection of stance on social media is an important task for public health applications, especially in the context of health crises. Unfortunately, existing models are typically trained on English corpora. Considering the benefits of extending research to other widely spoken languages, the goal of this study is to develop stance detection models for social media posts in Spanish.

MATERIALS AND METHODS: A corpus of 6170 tweets about COVID-19 vaccination, posted between March 1, 2020 and January 4, 2022, was manually annotated by native speakers. Traditional predictive models were compared with deep learning models to ascertain a baseline performance for the detection of stance in Spanish tweets. The evaluation focused on the ability of multilingual and language-specific embeddings to contextualize the topic of those short texts adequately.

RESULTS: The BERT-Multi+BiLSTM combination yielded the best results (macroaveraged F1 and Matthews correlation coefficient scores of 0.86 and 0.79, respectively; interpolated area under the receiver operating curve [AUC] of 0.95 for tweets against vaccination and 0.85 in favor of vaccination and a score of 0.97 for tweets containing no stance information), closely followed by the BETO+BiLSTM and RoBERTa BNE-LSTM Spanish models and the term frequency-inverse document frequency+SVM model (average AUC decrease of 0.01). The main differentiating factor among these models was the ability to predict tweets against vaccination.

DISCUSSION: The BERT Multi+BILSTM model outperformed the other models in terms of per class prediction capacity. The main assumption is that language-specific embeddings do not outperform multilingual embeddings or TF-IDF features because of the context of the topic. The inherent context of BERT or RoBERTa embeddings is general. So, these embeddings are not familiar with the slang commonly used on Twitter and, more specifically, during the pandemic.

CONCLUSION: The best performing model detects tweet stance with performance high enough to ensure its usefulness for public health applications, namely awareness campaigns, misinformation detection and other early intervention and prevention actions seeking to improve an individual's well-being based on autoreported experiences and opinions. The dataset and code of the study are available on GitHub.

PMID:40008184 | PMC:PMC11854073 | DOI:10.1093/jamiaopen/ooaf007

Front Public Health. 2025 Feb 11;13:1505825. doi: 10.3389/fpubh.2025.1505825. eCollection 2025.

ABSTRACT

INTRODUCTION: Parents' negative parenting style is an important cause of anxiety, depression, and suicide among university students. Given the widespread use of social media, microblogs offer a new and promising way for non-invasive, large-scale assessment of parenting styles of students' parents.

METHODS: In this study, we have two main objectives: (1) investigating the correlation between students' microblog discourses and parents' parenting styles and (2) devising a method to predict students' parenting styles from their microblog discourses. We analyzed 111,258 posts from 575 university students using frequency analysis to examine differences in the usage of topical and emotional word across different parenting styles. Informed by these insights, we developed an effective parenting style assessment method, including a correlation injection module.

RESULTS: Experimental results on the 575 students show that our method outperforms all the baseline NLP methods (including ChatGPT-4), achieving good assessment performance by reducing MSE by 14% to 0.12.

DISCUSSION: Our study provides a pioneering microblog-based parenting style assessment tool and constructs a dataset, merging insights from psychology and computational science. On the one hand, our study advances the understanding of how parenting styles are reflected in the linguistic and emotional expressions of students on microblogs. On the other hand, our study provides an assisting tool that could be used by healthcare institutions to identify students' parenting styles. It facilitates the identification of suicide risk factors among microblog student users, and enables timely interventions to prevent suicides, which enhances human wellbeing and saves lives.

PMID:40008146 | PMC:PMC11850274 | DOI:10.3389/fpubh.2025.1505825