Sci Rep. 2025 May 29;15(1):18858. doi: 10.1038/s41598-025-03208-z.

ABSTRACT

In the era of digitization, online digital advertising is one of the best techniques for modern marketing. This makes advertisers rely heavily on accurate user interest and behavior modelling to deliver precise advertisement impressions and increase click-through rates. The classic approach to model user interests has often required the use of predefined feature sets which are typically stagnant and not representative of temporal changes and trends in user behavior. While recent advances in deep learning offer potential solutions to these obstacles, many existing approaches fail to address the sequential nature of user interactions. In this paper, we propose an optimized Bi-Directional Long Short-Term Memory (Bi-LSTM) based user interest modeling approach together with an Updated version of Parrot Optimizer (UPO) to enhance performance. It treats the user activity as a temporal sequence which well fits the changing nature of user interest and preferences over time. The proposed approach is evaluated on two important tasks: predicting the probability that a user will click on an ad and predicting the probability that a user will click on a particular type of ad campaign. Simulation results demonstrate that the proposed method provides superior results than the static set-based approaches and achieves significant improvements on both user ad responses predictions and user ad clicks at the campaign level. The research also enhances the efficiency of user interest modeling with significant implications for online advertising, recommendation systems, and personalized marketing.

PMID:40442252 | DOI:10.1038/s41598-025-03208-z

BMC Med Educ. 2025 May 29;25(1):801. doi: 10.1186/s12909-025-07396-0.

ABSTRACT

BACKGROUND: Artificial intelligence (AI) has gained significant attention and progress in various scientific fields, especially medicine. Since its introduction in the 1950s, AI has advanced remarkably, supporting innovations like diagnostic tools and healthcare technologies. Despite these developments, challenges such as ethical concerns and limited integration in regions like Syria emphasize the importance of increasing awareness and conducting more targeted studies.

METHODS: A cross-sectional study was conducted to evaluate medical students' preparedness and readiness to use AI technologies in the medical field using the Medical Artificial Intelligence Readiness Scale for Medical Students (MAIRS_MS). The scale comprises 22 items divided into 4 domains: ethics, vision, ability, and cognition, with responses rated on a five-point Likert scale, higher scores indicate greater readiness. Data were collected through electronic and paper questionnaires distributed over a period of 20 days.

RESULTS: The study included 564 medical students from various Syrian universities, of whom 77.8% demonstrated awareness of AI in the medical field. Significant differences in AI awareness were observed based on academic GPA (p = 0.035) and income level (p = 0.016), with higher awareness among students with higher GPA and income levels. Statistically significant differences were found between students aware of AI and those unaware, as well as between students with experience using AI and those without, across all domains of readiness, including cognition (t = -10.319, p < 0.001), ability (t = -11.519, p < 0.001), vision (t = -6.387, p < 0.001), ethics (t = -7.821, p < 0.001), and the overall readiness score (t = -11.354, p < 0.001).

CONCLUSION: Integrating AI into medical education is essential for advancing healthcare in developing countries like Syria. Providing incentives and fostering a culture of continuous learning will equip medical students to leverage AI's benefits while mitigating its drawbacks.

PMID:40442679 | DOI:10.1186/s12909-025-07396-0

BMC Public Health. 2025 May 29;25(1):1977. doi: 10.1186/s12889-025-23156-7.

ABSTRACT

BACKGROUND: Environmental factors have been identified as significant risk factors for scrub typhus. However, the impact of inorganic compounds such as greenhouse gases and air pollutants on the incidence of scrub typhus has not been evaluated.

METHODS: Our study investigated the correlation between greenhouse gases, air pollutants from the global atmospheric emissions database (2005-2018), and reported cases of scrub typhus from the Public Health Science Data Center. First, an early warning method was applied to estimate the epidemic threshold and the grading intensity threshold. Second, four statistical methods were used to assess the correlation and lag effects across different age groups and epidemic periods. Deep learning algorithms were employed to evaluate the predictive effect of environmental factors on the incidence of scrub typhus.

RESULTS: Using the Moving Epidemic Method (MEM) and Treed Distributed Lag Non-Linear Model (TDLNM), we found that the period from April to September is the epidemic season for scrub typhus in China. During this period, BC, CH4, NH3 and PM10 all reach key windows during their respective early warning lag periods. Interaction effects showed that increased CO exposure during the 0-2-month period led to an increased magnitude of the PM10 effect during the 3-7-month period. The Quantile-based G Computation (qgcomp) model revealed age-specific differences in susceptibility to environmental factors. In the Bayesian Kernel Machine Regression (BKMR) model, we identified NOx (RRmax (95% CI) = 103.14 (70.40, 135.87)) and NMVOC as the risk environmental factors for young adults, while CH4 (RRmax (95% CI) = 20.94 (9.26, 32.63)) was significantly associated with scrub typhus incidence in younger populations. For the elderly, N2O and NOx (RRmax (95% CI) = 30.23 (13.78, 46.68)) were identified as susceptibility factors for scrub typhus. The Weighted Quantile Sum (WQS) model revealed a significant risk effect of NOx on scrub typhus during periods of low risk, which are often overlooked (OR (95% CI) = 0.40 (0.23, 0.58)). During periods of medium to high risk, Convolutional Neural Networks (CNN) showed that environmental factors performed well in predicting the incidence of scrub typhus.

CONCLUSIONS: We found that most greenhouse gases and air pollutants increase the risk of contracting scrub typhus, mainly driven by CH4, NOx, and NMVOC. Among these, the primary high-level pollutants have long-term lag effects during the epidemic period. The correlation between environmental factors and scrub typhus incidence varies significantly across different age groups and risk periods. Among them, middle-aged and young individuals are more susceptible to the effects of exposure to mixed air pollutants. CNN algorithm can help develop a comprehensive early warning system for scrub typhus. These findings may have important implications for guiding effective public health interventions in the future. The primary interventions should focus on controlling greenhouse gas emissions and reducing air pollutants, which can, in turn, be used to support infectious disease monitoring systems through environmental monitoring. Moreover, given the cross-sectional approach of our study, these findings need to be confirmed through additional cohort studies.

PMID:40442614 | DOI:10.1186/s12889-025-23156-7

Abdom Radiol (NY). 2025 May 29. doi: 10.1007/s00261-025-04991-z. Online ahead of print.

ABSTRACT

OBJECTIVE: The diagnosis of liver fibrosis is usually based on histopathological examination of liver puncture specimens. Although liver puncture is accurate, it has invasive risks and high economic costs, which are difficult for some patients to accept. Therefore, this study uses deep learning technology to build a liver fibrosis diagnosis model to achieve non-invasive staging of liver fibrosis, avoid complications, and reduce costs.

METHODS: This study uses ultrasound examination to obtain pure liver parenchyma image section data. With the consent of the patient, combined with the results of percutaneous liver puncture biopsy, the degree of liver fibrosis indicated by ultrasound examination data is judged. The concept of Fibrosis Contrast Layer (FCL) is creatively introduced in our experimental method, which can help our model more keenly capture the significant differences in the characteristics of liver fibrosis of various grades. Finally, through label fusion (LF), the characteristics of liver specimens of the same fibrosis stage are abstracted and fused to improve the accuracy and stability of the diagnostic model.

RESULTS: Experimental evaluation demonstrated that our model achieved an accuracy of 85.6%, outperforming baseline models such as ResNet (81.9%), InceptionNet (80.9%), and VGG (80.8%). Even under a small-sample condition (30% data), the model maintained an accuracy of 84.8%, significantly outperforming traditional deep-learning models exhibiting sharp performance declines.

CONCLUSION: The training results show that in the whole sample data set and 30% small sample data set training environments, the FCLLF model's test performance results are better than those of traditional deep learning models such as VGG, ResNet, and InceptionNet. The performance of the FCLLF model is more stable, especially in the small sample data set environment. Our proposed FCLLF model effectively improves the accuracy and stability of liver fibrosis staging using non-invasive ultrasound imaging.

PMID:40442504 | DOI:10.1007/s00261-025-04991-z

Pediatr Radiol. 2025 May 29. doi: 10.1007/s00247-025-06266-7. Online ahead of print.

ABSTRACT

BACKGROUND: Cardiovascular magnetic resonance (CMR) cine imaging is the gold standard for assessing ventricular volumes and function. It typically requires two-dimensional (2D) bSSFP sequences and multiple breath-holds, which can be challenging for patients with limited breath-holding capacity. Three-dimensional (3D) cardiovascular magnetic resonance angiography (MRA) usually suffers from lengthy acquisition. Free-running 3D cine imaging with deep learning (DL) reconstruction offers a potential solution by acquiring both cine and angiography simultaneously.

OBJECTIVE: To evaluate the efficiency and accuracy of a ferumoxytol-enhanced 3D cine imaging MR sequence combined with DL reconstruction and Heart-NAV technology in patients with congenital heart disease.

MATERIALS AND METHODS: This Institutional Review Board approved this prospective study that compared (i) functional and volumetric measurements between 3 and 2D cine images; (ii) contrast-to-noise ratio (CNR) between deep-learning (DL) and compressed sensing (CS)-reconstructed 3D cine images; and (iii) cross-sectional area (CSA) measurements between DL-reconstructed 3D cine images and the clinical 3D MRA images acquired using the bSSFP sequence. Paired t-tests were used to compare group measurements, and Bland-Altman analysis assessed agreement in CSA and volumetric data.

RESULTS: Sixteen patients (seven males; median age 6 years) were recruited. 3D cine imaging showed slightly larger right ventricular (RV) volumes and lower RV ejection fraction (EF) compared to 2D cine, with a significant difference only in RV end-systolic volume (P = 0.02). Left ventricular (LV) volumes and EF were slightly higher, and LV mass was lower, without significant differences (P ≥ 0.05). DL-reconstructed 3D cine images showed significantly higher CNR in all pulmonary veins than CS-reconstructed 3D cine images (all P < 0.05).

CONCLUSION: Highly accelerated free-running 3D cine imaging with DL reconstruction shortens acquisition times and provides comparable volumetric measurements to 2D cine, and comparable CSA to clinical 3D MRA.

PMID:40442341 | DOI:10.1007/s00247-025-06266-7

Sci Rep. 2025 May 29;15(1):18783. doi: 10.1038/s41598-025-03778-y.

ABSTRACT

Accurate classification of midpalatal suture maturation stages is critical for orthodontic diagnosis, treatment planning, and the assessment of maxillary growth. Cone Beam Computed Tomography (CBCT) imaging offers detailed insights into this craniofacial structure but poses unique challenges for deep learning image recognition model design due to its high dimensionality, noise artifacts, and variability in image quality. To address these challenges, we propose a novel technique that highlights key image features through a simple filtering process to improve image clarity prior to analysis, thereby enhancing the learning process and better aligning with the distribution of the input data domain. Our preprocessing steps include region-of-interest extraction, followed by high-pass and Sobel filtering for emphasis of low-level features. The feature extraction integrates Convolutional Neural Networks (CNN) architectures, such as EfficientNet and ResNet18, alongside our novel Multi-Filter Convolutional Residual Attention Network (MFCRAN) enhanced with Discrete Cosine Transform (DCT) layers. Moreover, to better capture the inherent order within the data classes, we augment the supervised training process with a ranking loss by attending to the relationship within the label domain. Furthermore, to adhere to diagnostic constraints while training the model, we introduce a tailored data augmentation strategy to improve classification accuracy and robustness. In order to validate our method, we employed a k-fold cross-validation protocol on a private dataset comprising 618 CBCT images, annotated into five stages (A, B, C, D, and E) by expert evaluators. The experimental results demonstrate the effectiveness of our proposed approach, achieving the highest classification accuracy of 79.02%, significantly outperforming competing architectures, which achieved accuracies ranging from 71.87 to 78.05%. This work introduces a novel and fully automated framework for midpalatal suture maturation classification, marking a substantial advancement in orthodontic diagnostics and treatment planning.

PMID:40442312 | DOI:10.1038/s41598-025-03778-y

Sci Rep. 2025 May 29;15(1):18888. doi: 10.1038/s41598-025-03254-7.

ABSTRACT

This study aimed to determine whether trigeminal neuralgia can be diagnosed using convolutional neural networks (CNNs) based on plain X-ray skull images. A labeled dataset of 166 skull images from patients aged over 16 years with trigeminal neuralgia was compiled, alongside a control dataset of 498 images from patients with unruptured intracranial aneurysms. The images were randomly partitioned into training, validation, and test datasets in a 6:2:2 ratio. Classifier performance was assessed using accuracy and the area under the receiver operating characteristic (AUROC) curve. Gradient-weighted class activation mapping was applied to identify regions of interest. External validation was conducted using a dataset obtained from another institution. The CNN achieved an overall accuracy of 87.2%, with sensitivity and specificity of 0.72 and 0.91, respectively, and an AUROC of 0.90 on the test dataset. In most cases, the sphenoid body and clivus were identified as key areas for predicting trigeminal neuralgia. Validation on the external dataset yielded an accuracy of 71.0%, highlighting the potential of deep learning-based models in distinguishing X-ray skull images of patients with trigeminal neuralgia from those of control individuals. Our preliminary results suggest that plain x-ray can be potentially used as an adjunct to conventional MRI, ideally with CISS sequences, to aid in the clinical diagnosis of TN. Further refinement could establish this approach as a valuable screening tool.

PMID:40442191 | DOI:10.1038/s41598-025-03254-7

Sci Rep. 2025 May 29;15(1):18807. doi: 10.1038/s41598-025-00628-9.

ABSTRACT

Vegetation fires are most common in Southeast Asian (SEA) countries, causing biodiversity loss, habitat destruction, and air pollution. Accurately predicting fire occurrences in SEA remains challenging due to its complex spatiotemporal dynamics. Improved fire predictions enable timely interventions, helping to control and mitigate fires. In this study, we utilize Visible Infrared Imaging Radiometer Suite (VIIRS) satellite-derived fire data alongside six machine learning (ML) and deep learning (DL) models, Simple Persistence, Multi-Layer Perceptron (MLP), Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), CNN-Long Short-Term Memory (CNN-LSTM), and Convolutional Long Short-Term Memory (ConvLSTM) to determine the most effective fire prediction model. We evaluated model performance using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and R2 (coefficient of determination). Our results indicate that the CNN performs best in regions with strong spatial dependencies, such as Brunei, Indonesia, Malaysia, the Philippines, Timor-Leste, and Thailand. Conversely, the ConvLSTM excels in countries with complex spatiotemporal dynamics, like Laos, Myanmar, and Vietnam. The CNN-LSTM hybrid model also performed well in Cambodia, suggesting a need for a balanced approach in areas requiring both spatial and temporal feature extraction. Furthermore, simpler models, such as Simple Persistence and MLP, showed limitations in capturing dynamic patterns and temporal dependencies. Our findings highlight the importance of evaluating various ML and DL models before integrating them into any decision support systems (DSS) for fire management studies. By tailoring models to specific regional fire data, prediction accuracy and responsiveness can be enhanced, ultimately improving fire risk management in Southeast Asia and beyond.

PMID:40442135 | DOI:10.1038/s41598-025-00628-9

Nat Commun. 2025 May 29;16(1):4994. doi: 10.1038/s41467-025-60333-z.

ABSTRACT

Recent advancements in single-cell technologies have enabled comprehensive characterization of cellular states through transcriptomic, epigenomic, and proteomic profiling at single-cell resolution. These technologies have significantly deepened our understanding of cell functions and disease mechanisms from various omics perspectives. As these technologies evolve rapidly and data resources expand, there is a growing need for computational methods that can integrate information from different modalities to facilitate joint analysis of single-cell multi-omics data. However, integrating single-cell omics datasets presents unique challenges due to varied feature correlations and technology-specific limitations. To address these challenges, we introduce scMODAL, a deep learning framework tailored for single-cell multi-omics data alignment using feature links. scMODAL integrates datasets with limited known positively correlated features, leveraging neural networks and generative adversarial networks to align cell embeddings and preserve feature topology. Our experiments demonstrate scMODAL's effectiveness in removing unwanted variation, preserving biological information, and accurately identifying cell subpopulations across diverse datasets. scMODAL not only advances integration tasks but also supports downstream analyses such as feature imputation and feature relationship inference, offering a robust solution for advancing single-cell multi-omics research.

PMID:40442129 | DOI:10.1038/s41467-025-60333-z

J Chem Inf Model. 2025 May 29. doi: 10.1021/acs.jcim.5c00575. Online ahead of print.

ABSTRACT

Dysfunctions of the dopamine D2 and D3 receptors (D2 and D3) are implicated in neuropsychiatric conditions such as Parkinson's disease, schizophrenia, and substance use disorders (SUDs). Evidence indicates that D3-selective ligands can effectively modulate reward pathways, offering potential in treating SUDs with reduced side effects. However, the high homology between D2 and D3 presents challenges in developing subtype-selective ligands, crucial for elucidating receptor-specific functions and developing targeted therapeutics. Here, to facilitate selective ligand discovery, we leveraged ligand-based quantitative structure-activity relationship (QSAR) modeling approaches to predict binding affinity at D2 and D3, as well as ligand selectivity for D3. We first queried training data from the ChEMBL database and performed a systematic curation process to enhance the data quality. We then developed QSAR models using eXtreme Gradient Boosting, random forest, and deep neural network (DNN) algorithms, with DNN benefiting from a novel hyperparameter optimization protocol. All models exhibited strong predictive performance, with DNN-based models slightly but consistently outperforming the tree-based models. Integrating predictions from all algorithms into a consensus metric further improved the accuracy and robustness. Notably, our selectivity models outperformed the affinity models, likely due to noise cancellation achieved by subtracting the binding affinities of the two receptors. The Shapley Additive explanations analysis revealed key pharmacophoric and physicochemical features critical for receptor affinity and selectivity, while molecular docking of representative D3-selective compounds highlighted the structural basis of D3 selectivity. These findings provide a robust framework for modeling QSARs at D2 and D3, advancing the rational design of targeted therapeutics for these receptors.

PMID:40442044 | DOI:10.1021/acs.jcim.5c00575

Obstet Gynecol Sci. 2025 May 29. doi: 10.5468/ogs.25041. Online ahead of print.

ABSTRACT

Artificial intelligence (AI) and machine learning (ML) are transforming cervical cancer research and offering advancements in diagnosis, prognosis, screening, and treatment. This review explores ML applications with particular emphasis on prediction models. A comprehensive literature search identified studies using ML for survival prediction, risk assessment, and treatment optimization. ML-driven prognostic models integrate clinical, histopathological, and genomic data to improve survival prediction and patient stratification. Screening methods, including deep-learning-based cytology analysis and HPV detection, enhance accuracy and efficiency. ML-driven imaging techniques facilitate early and precise cancer diagnosis, whereas risk prediction models assess susceptibility based on demographic and genetic factors. AI also optimizes treatment planning by predicting therapeutic responses and guiding personalized interventions. Despite significant progress, challenges remain regarding data availability, model interpretability, and clinical implementation. Standardized datasets, external validation, and cross-disciplinary collaborations are crucial for implementing ML innovations in clinical settings. Subsequent investigations should prioritize joint initiatives among data scientists, healthcare providers, and health authorities to translate AI innovations into real-world applications and to enhance the impact of ML on cervical cancer care. By synthesizing recent developments, this review highlights the potential of ML to improve clinical outcomes and shaping the future of cervical cancer management.

PMID:40441737 | DOI:10.5468/ogs.25041

Med Image Anal. 2025 May 23;104:103633. doi: 10.1016/j.media.2025.103633. Online ahead of print.

ABSTRACT

Invasive breast cancer (IBC) is a prevalent malignant tumor in women, and precise grading plays a pivotal role in ensuring effective treatment and enhancing survival rates. However, accurately grading IBC presents a significant challenge due to its heterogeneous nature and the need to harness the complementary information from multiple nuclei sources in histopathology images. To tackle this critical problem, we introduce a novel multi-source hybrid feature fusion network named MSFusion. This network incorporates two types of hybrid features: deep learning features extracted by a novel Swin Transformer-based multi-branch network called MSwinT, and traditional handcrafted features that capture the morphological characteristics of multi-source nuclei. The primary branch of MSwinT captures the overall characteristics of the original images, while multiple auxiliary branches focus on identifying morphological features from diverse sources of nuclei, including tumor, mitotic, tubular, and epithelial nuclei. At each of the four stages for the branches in MSwinT, a functional KDC (key diagnostic components) fusion block with channel and spatial attentions is proposed to integrate the features extracted by all the branches. Ultimately, we synthesize the multi-source hybrid deep learning features and handcrafted features to improve the accuracy of IBC diagnosis and grading. Our multi-branch MSFusion network is rigorously evaluated on three distinct datasets, including two private clinical datasets (Qilu dataset and QDUH&SHSU dataset) as well as a publicly available Databiox dataset. The experimental results consistently demonstrate that our proposed MSFusion model outperforms the state-of-the-art methods. Specifically, the AUC for the Qilu dataset and QDUH&SHSU dataset are 81.3% and 90.2%, respectively, while the public Databiox dataset yields an AUC of 82.1%.

PMID:40441045 | DOI:10.1016/j.media.2025.103633

EBioMedicine. 2025 May 28;116:105750. doi: 10.1016/j.ebiom.2025.105750. Online ahead of print.

ABSTRACT

BACKGROUND: Ductal Carcinoma In Situ (DCIS) can progress to ipsilateral invasive breast cancer (IBC) but over 75% of DCIS lesions do not progress if untreated. Currently, DCIS that might progress to IBC cannot reliably be identified. Therefore, most patients with DCIS undergo treatment resembling IBC. To facilitate identification of low-risk DCIS, we developed deep learning models using histology whole-slide images (WSIs) and clinico-pathological data.

METHODS: We predicted invasive recurrence in patients with primary, pure DCIS treated with breast-conserving surgery using clinical Cox proportional hazards models and deep learning. Deep learning models were trained end-to-end with only WSIs or in combination with clinical data (integrative). We employed nested k-fold cross-validation (k = 5) on a Dutch multicentre dataset (n = 558). Models were also tested on the UK-based Sloane dataset (n = 94).

FINDINGS: Evaluated over 20 years on the Dutch dataset, deep learning models using only WSIs effectively stratified patients into low-risk (no recurrence) and high-risk (invasive recurrence) groups (negative predictive value (NPV) = 0.79 (95% CI: 0.74-0.83); hazard ratio (HR) = 4.48 (95% CI: 3.41-5.88, p < 0.0001); area under the receiver operating characteristic curve (AUC) = 0.75 (95% CI: 0.70-0.79)). Integrative models achieved similar results with slightly enhanced hazard ratios compared to the image-only models (NPV = 0.77 (95% CI 0.73-0.82); HR = 4.85 (95% CI 3.65-6.45, p < 0.0001); AUC = 0.75 (95% CI 0.7-0.79)). In contrast, clinical models were borderline significant (NPV = 0.64 (95% CI 0.59-0.69); HR = 1.37 (95% CI 1.03-1.81, p = 0.041); AUC = 0.57 (95% CI 0.52-0.62)). Furthermore, external validation of the models was unsuccessful, limited by the small size and low number of cases (22/94) in our external dataset, WSI quality, as well as the lack of well-annotated datasets that allow robust validation.

INTERPRETATION: Deep learning models using routinely processed WSIs hold promise for DCIS risk stratification, while the benefits of integrating clinical data merit further investigation. Obtaining a larger, high-quality external multicentre dataset would be highly valuable, as successful generalisation of these models could demonstrate their potential to reduce overtreatment in DCIS by enabling active surveillance for women at low risk.

FUNDING: Cancer Research UK, the Dutch Cancer Society (KWF), and the Dutch Ministry of Health, Welfare and Sport.

PMID:40440915 | DOI:10.1016/j.ebiom.2025.105750

Forensic Sci Int. 2025 May 21;372:112511. doi: 10.1016/j.forsciint.2025.112511. Online ahead of print.

ABSTRACT

Accurate adult age estimation (AAE) is critical for forensic and anthropological applications, yet traditional methods relying on bone mineral density (BMD) face significant challenges due to biological variability and methodological limitations. This study aims to develop an end-to-end Deep Learning (DL) based pipeline for automated AAE using BMD from proximal femoral CT scans. The main objectives are to construct a large-scale dataset of 5151 CT scans from real-world clinical and cadaver cohorts, fine-tune the Segment Anything Model (SAM) for accurate femoral bone segmentation, and evaluate multiple convolutional neural networks (CNNs) for precise age estimation based on segmented BMD data. Model performance was assessed through cross-validation, internal clinical testing, and external post-mortem validation. SAM achieved excellent segmentation performance with a Dice coefficient of 0.928 and an average intersection over union (mIoU) of 0.869. The CNN models achieved an average mean absolute error (MAE) of 5.20 years in cross-validation (male: 5.72; female: 4.51), which improved to 4.98 years in the independent clinical test set (male: 5.32; female: 4.56). External validation on the post-mortem dataset revealed an MAE of 6.91 years, with 6.97 for males and 6.69 for females. Ensemble learning further improved accuracy, reducing MAE to 4.78 years (male: 5.12; female: 4.35) in the internal test set, and 6.58 years (male: 6.64; female: 6.37) in the external validation set. These findings highlight the feasibility of dl-driven AAE and its potential for forensic applications, offering a fully automated framework for robust age estimation.

PMID:40440868 | DOI:10.1016/j.forsciint.2025.112511

Comput Methods Programs Biomed. 2025 May 27;269:108844. doi: 10.1016/j.cmpb.2025.108844. Online ahead of print.

ABSTRACT

BACKGROUND AND OBJECTIVE: Despite the profound advancements that deep learning models have achieved across a multitude of domains, their propensity to learn spurious correlations significantly impedes their applicability to tasks necessitating causal and counterfactual reasoning.

METHODS: In this paper, we propose a Bidirectional Neural Network, which innovatively consolidates forward causal reasoning with inverse counterfactual reasoning into a cohesive framework. This integration is facilitated through the implementation of multi-stacked affine coupling layers, which ensure the network's invertibility, thereby enabling bidirectional reasoning capabilities within a singular architectural construct. To augment the network's trainability and to ensure the bidirectional differentiability of the parameters, we introduce an orthogonal weight normalization technique. Additionally, the counterfactual reasoning capacity of the Bidirectional Neural Network is embedded within the policy function of reinforcement learning, thereby effectively addressing the challenges associated with reward sparsity in the blood glucose control scenario.

RESULTS: We evaluate our framework on two pivotal tasks: causal-based blood glucose forecasting and counterfactual-based blood glucose control. The empirical results affirm that our model not only exemplifies enhanced generalization in causal reasoning but also significantly surpasses comparative models in handling out-of-distribution data. Furthermore, in blood glucose control tasks, the integration of counterfactual reasoning markedly improves decision efficacy, sample efficiency, and convergence velocity.

CONCLUSION: It is our expectation that the Bidirectional Neural Network will pave novel pathways in the exploration of causal and counterfactual reasoning, thus providing groundbreaking methods for complex decision-making processes. Code is available at https://github.com/HITshenrj/BNN.

PMID:40440769 | DOI:10.1016/j.cmpb.2025.108844

Percept Mot Skills. 2025 May 29:315125251347006. doi: 10.1177/00315125251347006. Online ahead of print.

ABSTRACT

Frequency-following responses (FFRs) are neural signals that reflect the brain's encoding of acoustic characteristics, such as speech intonation. While traditional machine learning models have been used to classify FFRs elicited under various conditions, the potential of deep learning models in FFR research remains underexplored. This study investigated the efficacy of a three-layer artificial neural network (ANN) in detecting the presence or absence of FFRs elicited by a rising intonation of the English vowel /i/. The ANN was trained and tested on FFR recordings, using F0 estimates derived from the spectral domain as input data. Model performance was evaluated by systematically varying the number of inputs, hidden neurons, and the number of sweeps included in the recordings. The prediction accuracy of the ANN was significantly influenced by the number of inputs, hidden neurons, and sweeps. Optimal configurations included 6-8 inputs and 4-6 hidden neurons, achieving a prediction accuracy of approximately 84% when the signal-to-noise ratio was enhanced by including 100 or more sweeps. These results provide a foundation for future applications in auditory processing assessments and clinical diagnostics.

PMID:40440687 | DOI:10.1177/00315125251347006

Int J Surg. 2025 May 29. doi: 10.1097/JS9.0000000000002526. Online ahead of print.

ABSTRACT

BACKGROUND: Currently, there remains a critical need for reliable tools to accurately predict post-surgical outcomes in non-small cell lung cancer (NSCLC) patients in clinical practice. We aimed to develop and validate a deep learning-based model utilizing histopathological slides to accurately predict post-surgical outcomes in NSCLC patients.

METHODS: In this study, we analyzed histopathological slides and comprehensive clinical data from 337 Local-NSCLC patients for model development, and further validated the model using an independent cohort of 554 NSCLC patients from The Cancer Genome Atlas (TCGA) database. Utilizing the advanced Res2Net deep learning architecture, we developed and optimized a novel Surgical Prognosis Prediction Score (Sr-PPS) system.

RESULTS: The Sr-PPS model demonstrated significantly enhanced predictive accuracy for both disease-free survival (DFS) and overall survival (OS) in NSCLC patients. Multivariate Cox regression analysis validated Sr-PPS as a robust independent predictor of post-surgical outcomes in NSCLC patients. Patients with low Sr-PPS scores exhibited enhanced anti-tumor immune microenvironment characteristics, characterized by significant upregulation of immune activation pathways (particularly T-cell migration and B-cell receptor signaling), coupled with marked downregulation of oncogenic pathways, including insulin-like growth factor receptor signaling and STAT protein phosphorylation. Further genomic analyses revealed significant associations between Sr-PPS scores and mutations in key oncogenic driver genes, including CTNND2, PRRX1, and ALK.

CONCLUSIONS: Our deep learning-based Sr-PPS model not only demonstrates robust predictive capability for post-surgical outcomes in NSCLC patients but also elucidates underlying molecular mechanisms, thereby providing a valuable framework for personalized treatment stratification.

PMID:40440686 | DOI:10.1097/JS9.0000000000002526

JMIR Med Inform. 2025 May 29;13:e67859. doi: 10.2196/67859.

ABSTRACT

BACKGROUND: Leptospirosis, a zoonotic disease caused by Leptospira bacteria, continues to pose significant public health risks, particularly in tropical and subtropical regions.

OBJECTIVE: This systematic review aimed to evaluate the application of machine learning (ML) and deep learning (DL) techniques in predicting and diagnosing leptospirosis, focusing on the most used algorithms, validation methods, data types, and performance metrics.

METHODS: Using Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines, Checklist for Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies (CHARMS), and Prediction model Risk of Bias Assessment Tool (PROBAST) tools, we conducted a comprehensive review of studies applying ML and DL models for leptospirosis detection and prediction, examining algorithm performance, data sources, and validation approaches.

RESULTS: Out of a total of 374 articles screened, 17 studies were included in the qualitative synthesis, representing approximately 4.5% of the initial pool. The review identified frequent use of algorithms such as support vector machines, artificial neural networks, decision trees, and convolutional neural networks (CNNs). Among the included studies, 88% (15/17) used traditional ML methods, and 24% (4/17) used DL techniques. Several models demonstrated high predictive performance, with reported accuracy rates ranging from 80% to 98%, notably with the U-Net CNN achieving 98.02% accuracy. However, public datasets were underused, with only 35% (6/17) of studies incorporating publicly available data sources; the majority (65%, 11/17) relied primarily on private datasets from hospitals, clinical records, or regional surveillance systems.

CONCLUSIONS: ML and DL techniques demonstrate potential for improving leptospirosis prediction and diagnosis, but future research should focus on using larger, more diverse datasets, adopting transfer learning strategies, and integrating advanced ensemble and validation techniques to strengthen model accuracy and generalization.

PMID:40440642 | DOI:10.2196/67859

J Chem Inf Model. 2025 May 29. doi: 10.1021/acs.jcim.5c00489. Online ahead of print.

ABSTRACT

AlphaFold2 and other deep learning tools represent the state of the art for protein structure prediction; however, they are still limited in their ability to model multiple protein conformations. Since the function of many proteins depends on their ability to assume different stable conformational states, different approaches are required to access these alternative conformations. G-protein-coupled receptors regulate intracellular signaling by assuming two main conformational states: an active state able to bind G-protein and an inactive state. Receptor activation is characterized by large conformational changes at the intracellular region, where the G-protein interacts, accompanied by more subtle structural rearrangements at the extracellular ligand-binding site. Retrospective studies have demonstrated that, for many receptors, the inactive state is the favored conformation generated by AlphaFold2 when the receptor is modeled alone, while active-state structures can only be modeled by introducing a conformational bias in the template information used for the prediction or by explicitly incorporating the binding of a ligand into the modeled system. This benchmarking study extends previous analyses, confirming the opportunities of deep learning tools for modeling G-protein complexed to the active state of receptor, while also revealing limitations in the modeling of allosteric effects, particularly the reduced accuracy of predictions at the receptor extracellular site, which may impact their applicability in structure-based drug design.

PMID:40440630 | DOI:10.1021/acs.jcim.5c00489

JMIR Form Res. 2025 May 29;9:e67311. doi: 10.2196/67311.

ABSTRACT

BACKGROUND: Deaths related to physical trauma impose a heavy burden on society, and the Abbreviated Injury Scale (AIS) is an important tool for injury research. AIS covers injuries to various parts of the human body and scores them based on the severity of the injury. In practical applications, the complex AIS coding rules require experts to encode by consulting patient medical records, which inevitably increases the difficulty, time, and cost of evaluation of patient and also puts higher demands on the workload of information collection and processing. In some cases, the sheer number of patients or the inability to access detailed medical records necessary for coding further complicates independent AIS codes.

OBJECTIVE: This study aims to use advanced deep learning techniques to predict AIS codes based on easily accessible diagnostic information of patients to improve the accuracy of trauma assessment.

METHODS: We used a dataset of patients with trauma (n=26,810) collected by the Chongqing Daping Hospital between October 2013 and June 2024. We mainly selected the patient's diagnostic information, injury description, cause of injury, injury region, injury types, and present illness history as the key feature inputs. We used a robust optimization Bidirectional Encoder Representations from Transformers (BERT) pretraining method to embed these features and constructed a prediction model based on BERT. This model aims to predict AIS codes and comprehensively evaluate its performance through a 5-fold cross-validation. We compared the BERT model with previous research results and current mainstream machine learning methods to verify its advantages in prediction tasks. In addition, we also conducted external validation of the model using 244 external data points from the Chongqing Emergency Center.

RESULTS: The BERT model proposed in this paper performs significantly better than the comparison model on independent test datasets with an accuracy of 0.8971, which surpassed the previous study by 10 % points. In addition, the area under the curve (AUC value of the BERT model is 0.9970, and the F1-score is 0.8434. In the external dataset, the accuracy, AUC, and F1-score results of the model are 0.7131, 0.8586, and 0.6801, respectively. These results indicate that our model has high generalization ability and prediction accuracy.

CONCLUSIONS: The BERT model we proposed is mainly based on diagnostic information to predict AIS codes, and its prediction accuracy is superior to previous investigations and current mainstream machine learning methods. It has a high generalization ability in external datasets.

PMID:40440586 | DOI:10.2196/67311