Brief Bioinform. 2025 May 1;26(3):bbaf210. doi: 10.1093/bib/bbaf210.

ABSTRACT

Spatial omics technologies, generating high-throughput and multimodal data, have necessitated the development of advanced data integration methods to facilitate comprehensive biological and clinical treatment discoveries. Based on the cross-attention concept, we developed an AI learning based toolchain called StereoMM, a graph based fusion model that can incorporate omics data such as gene expression, histological images, and spatial location. StereoMM uses an attention module for omics data interaction and a graph autoencoder to integrate spatial positions and omics data in a self-supervised manner. Applying StereoMM across various cancer types and platforms has demonstrated its robust capability. StereoMM outperforms competitors in identifying spatial regions reflecting tumour progression and shows promise in classifying colorectal cancer patients into deficient mismatch repair and proficient mismatch repair groups. The comprehensive inter-modal integration and efficiency of StereoMM enable researchers to construct spatial views of integrated multimodal features efficiently, advancing thorough tissue and patient characterization.

PMID:40407386 | DOI:10.1093/bib/bbaf210

Bioinformatics. 2025 May 23:btaf317. doi: 10.1093/bioinformatics/btaf317. Online ahead of print.

ABSTRACT

MOTIVATION: Modeling genetic perturbations and their effect on the transcriptome is a key area of pharmaceutical research. Due to the complexity of the transcriptome, there has been much excitement and development in deep learning (DL) because of its ability to model complex relationships. In particular, the transformer-based foundation model paradigm emerged as the gold-standard of predicting post-perturbation responses. However, understanding these increasingly complex models and evaluating their practical utility is lacking, along with simple but appropriate benchmarks to compare predictive methods.

RESULTS: Here, we present a simple baseline method that outperforms both state of the art (SOTA) in DL and other proposed simpler neural architectures, setting a necessary benchmark to evaluate in the field of post-perturbation prediction. We also elucidate the utility of foundation models for the task of post-perturbation prediction via generalizable fine-tuning experiments that can be translated to different applications of transformer-based foundation models to tasks of interest. Furthermore, we provide a corrected version of a popular dataset used for benchmarking perturbation prediction models. Our hope is that this work will properly contextualize further development of DL models in the perturbation space with necessary control procedures.

AVAILABILITY AND IMPLEMENTATION: All source code is available at: https://github.com/pfizer-opensource/perturb_seq. The DOI is 10.5281/zenodo.15352937.

CONTACT: daniel.wong@pfizer.com.

SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

PMID:40407144 | DOI:10.1093/bioinformatics/btaf317

Spine (Phila Pa 1976). 2025 May 23. doi: 10.1097/BRS.0000000000005381. Online ahead of print.

ABSTRACT

STUDY DESIGN: Retrospective image analysis study.

OBJECTIVE: To propose a novel classification system for adolescent idiopathic scoliosis (AIS) curvature using unsupervised machine learning and evaluate its reliability and clinical implications.

SUMMARY OF BACKGROUND DATA: Existing AIS classification systems, such as King and Lenke, have limitations in accurately describing curve variations, particularly long C-shaped curves or curves with distinct characteristics. Unsupervised machine learning offers an opportunity to refine classification and enhance clinical decision-making.

METHODS: A total of 1,156 AIS patients who underwent deformity correction surgery were analyzed. Standard posteroanterior radiographs were segmented using U-net algorithms. Contrastive clustering was employed for automatic grouping, with the number of clusters ranging from three to 10. Cluster quality was assessed using t-SNE and Silhouette scores. Clusters were defined based on consensus among spine surgeons. Interobserver reliability was evaluated using kappa coefficients.

RESULTS: Six clusters were identified, reflecting variations in structural curve location, single (C-shaped) versus double (S-shaped) curves, and thoracolumbar curve characteristics. Cluster reliability was moderate (kappa = 0.701-0.731). The silhouette score was 0.308, with t-SNE demonstrating distinct clustering patterns. The classification highlighted differences not captured by the Lenke classification, such as thoracic curves confined to the thoracic spine versus those extending to the lumbar spine.

CONCLUSION: Unsupervised machine learning successfully categorized AIS curvatures into six distinct clusters, revealing meaningful patterns such as unique variations in thoracic and lumbar curves. These findings could potentially inform surgical planning and prognostic assessments. However, further studies are needed to validate clinical applicability and improve clustering quality.

LEVEL OF EVIDENCE: 3.

PMID:40407029 | DOI:10.1097/BRS.0000000000005381

J Cardiovasc Electrophysiol. 2025 May 23. doi: 10.1111/jce.16733. Online ahead of print.

ABSTRACT

BACKGROUND: The long-term success rate of atrial fibrillation (AF) ablation remains a significant clinical challenge, particularly in patients with persistent atrial fibrillation (Persistent AF, PeAF). The recurrence risk in PeAF patients is influenced by various factors, which complicates the prediction of ablation outcomes. While clinical characteristics provide important references for risk assessment, the predictive accuracy of existing methods is limited and they fail to fully leverage the rich information contained in electrocardiogram (ECG) signals. Integrating clinical features with ECG signals holds promise for enhancing recurrence prediction accuracy and supporting personalized management.

METHODS: This study conducted a retrospective analysis of PeAF patients who underwent radiofrequency catheter ablation treatment between 2016 and 2019. A multimodal fusion framework based on a residual block network structure was proposed, integrating preprocedural AF rhythm 12-lead ECG signals, clinical scores, and baseline characteristics of the patients to construct a deep learning model for predicting the risk of postablation recurrence in PeAF patients. A fivefold cross-validation method was used to partition the data set for model training and testing.

RESULTS: The fusion model was evaluated on a cohort of 77 PeAF patients, achieving good predictive performance with an average AUC of 0.74, and a maximum of 0.82. It significantly outperformed traditional clinical scoring systems and single-modal models based solely on ECG signals. Additionally, the model demonstrated lower variance (0.08), reflecting its robustness and stability with small sample sizes.

CONCLUSION: This study innovatively combines AF rhythm ECG signals with clinical characteristics to construct a deep learning model for predicting the recurrence risk in PeAF patients after radiofrequency catheter ablation. The results show that this method effectively improves prediction performance and provides support for personalized clinical decision-making, with significant potential for clinical application.

PMID:40406972 | DOI:10.1111/jce.16733

Neuroradiol J. 2025 May 23:19714009251345108. doi: 10.1177/19714009251345108. Online ahead of print.

ABSTRACT

Background and purpose: This study focuses on an objective evaluation of a novel reconstruction algorithm-Deep Learning Image Reconstruction (DLIR)-ability to improve image quality and reduce radiation dose compared to the established standard of Adaptive Statistical Iterative Reconstruction-V (ASIR-V), in unenhanced head computed tomography (CT). Materials and methods: A retrospective analysis of 163 consecutive unenhanced head CTs was conducted. Image quality assessment was computed on the objective parameters of Signal-to-Noise Ratio (SNR) and Contrast-to-Noise Ratio (CNR), derived from 5 regions of interest (ROI). The evaluation of DLIR dose reduction abilities was based on the analysis of the PACS derived parameters of dose length product and computed tomography dose index volume (CTDIvol). Results: Following the application of rigorous criteria, the study comprised 35 patients. Significant image quality improvement was achieved with the implementation of DLIR, as evidenced by up to a 145% and 160% increase in SNR in supra- and infratentorial regions, respectively. CNR measurements further confirmed the superiority of DLIR over ASIR-V, with an increase of 171.5% in the supratentorial region and a 59.3% increase in the infratentorial region. Despite the signal improvement and noise reduction DLIR facilitated radiation dose reduction of up to 44% in CTDIvol. Conclusion: Implementation of DLIR in head CT scans enables significant image quality improvement and dose reduction abilities compared to standard ASIR-V. However, the dose reduction feature was proven insufficient to counteract the lack of gantry angulation in wide-detector scanners.

PMID:40406852 | DOI:10.1177/19714009251345108

Front Plant Sci. 2025 May 8;16:1615289. doi: 10.3389/fpls.2025.1615289. eCollection 2025.

NO ABSTRACT

PMID:40406730 | PMC:PMC12095160 | DOI:10.3389/fpls.2025.1615289

Front Neurol. 2025 Apr 24;16:1544571. doi: 10.3389/fneur.2025.1544571. eCollection 2025.

ABSTRACT

BACKGROUND: The traditional procedure of intracranial aneurysm (IA) diagnosis and evaluation in MRA is manually operated, which is time-consuming and labor-intensive. In this study, a deep learning model was established to diagnose and measure IA automatically based on the original MR images.

METHODS: A total of 1,014 IAs (from 852 patients) from hospital 1 were included and randomly divided into training, testing, and internal validation sets in a 7:2:1 ratio. Additionally, 315 patients (179 cases with IA and 136 cases without IA) from hospital 2 were used for independent validation. A deep learning model of MR 3DUnet was established for IA diagnosis and size measurement. The true positive (TP), false positive (FP), false negative (FN), recall, sensitivity, and specificity indices were used to evaluate the diagnosis performance of MR 3DUnet. The two-sample t-test was used to compare the size measurement results of MR 3DUnet and two radiologists. A p-value of < 0.05 was considered statistically significant.

RESULTS: The fully automatic model processed the original MRA data in 13.6 s and provided real-time results, including IA diagnosis and size measurement. For the IA diagnosis, in the training, testing, and internal validation sets, the recall rates were 0.80, 0.75, and 0.79, and the sensitivities were 0.82, 0.75, and 0.75, respectively. In the independent validation set, the recall rate, sensitivity, specificity, and AUC were 0.71, 0.74, 0.77, and 0.75, respectively. Subgroup analysis showed a recall rate of 0.74 for IA diagnosis based on DSA. For IA size measurement, no significant difference was found between our MR 3DUnet and the manual measurements of DSA or MRA.

CONCLUSION: In this study, a one-click, fully automatic deep learning model was developed for automatic IA diagnosis and size measurement based on 2D original images. It has the potential to significantly improve doctors' work efficiency and reduce patients' examination time, making it highly valuable in clinical practice.

PMID:40406704 | PMC:PMC12096847 | DOI:10.3389/fneur.2025.1544571

Proc SPIE Int Soc Opt Eng. 2025 Feb;13410:134100Q. doi: 10.1117/12.3047233. Epub 2025 Apr 2.

ABSTRACT

The Adolescent Brain Cognitive Development (ABCD) Study collects extensive neuroimaging data, including over 20,000 MRI sessions, to understand brain development in children. Ensuring high-quality MRI data is essential for accurate analysis, but manual Quality Control (QC) is impractical for large datasets due to time and resource constraints. We propose an automated QC method using an ensemble model that leverages metrics from FSQC and a novel deep learning model for brain shape analysis that uses cortical thickness, curvature, sulcal depth, and surface area as input features. The ensemble model achieved an accuracy of 76%, while our method achieved an accuracy of 72.62%, with balanced precision, recall, and F1 scores for both classes. This automated method promises to improve QC processes and accelerate the analysis of ABCD data.

PMID:40406668 | PMC:PMC12096335 | DOI:10.1117/12.3047233

Front Bioeng Biotechnol. 2025 May 8;13:1580502. doi: 10.3389/fbioe.2025.1580502. eCollection 2025.

ABSTRACT

OBJECTIVES: Accurate segmentation of craniomaxillofacial (CMF) structures and individual teeth is essential for advancing computer-assisted CMF surgery. This study developed CMF-ELSeg, a novel fully automatic multi-structure segmentation model based on deep ensemble learning.

METHODS: A total of 143 CMF computed tomography (CT) scans were retrospectively collected and manually annotated by experts for model training and validation. Three 3D U-Net-based deep learning models (V-Net, nnU-Net, and 3D UX-Net) were benchmarked. CMF-ELSeg employed a coarse-to-fine cascaded architecture and an ensemble approach to integrate the strengths of these models. Segmentation performance was evaluated using Dice score and Intersection over Union (IoU) by comparing model predictions to ground truth annotations. Clinical feasibility was assessed through qualitative and quantitative analyses.

RESULTS: In coarse segmentation of the upper skull, mandible, cervical vertebra, and pharyngeal cavity, 3D UX-Net and nnU-Net achieved Dice scores above 0.96 and IoU above 0.93. For fine segmentation and classification of individual teeth, the cascaded 3D UX-Net performed best. CMF-ELSeg improved Dice scores by 3%-5% over individual models for facial soft tissue, upper skull, mandible, cervical vertebra, and pharyngeal cavity segmentation, and maintained high accuracy Dice > 0.94 for most teeth. Clinical evaluation confirmed that CMF-ELSeg performed reliably in patients with skeletal malocclusion, fractures, and fibrous dysplasia.

CONCLUSION: CMF-ELSeg provides high-precision segmentation of CMF structures and teeth by leveraging multiple models, serving as a practical tool for clinical applications and enhancing patient-specific treatment planning in CMF surgery.

PMID:40406586 | PMC:PMC12094958 | DOI:10.3389/fbioe.2025.1580502

Flow Turbul Combust. 2025;115(1):3-27. doi: 10.1007/s10494-025-00642-x. Epub 2025 Mar 5.

ABSTRACT

This study presents novel drag reduction active-flow-control (AFC) strategies for a three-dimensional cylinder immersed in a flow at a Reynolds number based on freestream velocity and cylinder diameter of R e D = 3900 . The cylinder in this subcritical flow regime has been extensively studied in the literature and is considered a classic case of turbulent flow arising from a bluff body. The strategies presented are explored through the use of deep reinforcement learning. The cylinder is equipped with 10 independent zero-net-mass-flux jet pairs, distributed on the top and bottom surfaces, which define the AFC setup. The method is based on the coupling between a computational-fluid-dynamics solver and a multi-agent reinforcement-learning (MARL) framework using the proximal-policy-optimization algorithm. This work introduces a multi-stage training approach to expand the exploration space and enhance drag reduction stabilization. By accelerating training through the exploitation of local invariants with MARL, a drag reduction of approximately 9 % is achieved. The cooperative closed-loop strategy developed by the agents is sophisticated, as it utilizes a wide bandwidth of mass-flow-rate frequencies, which classical control methods are unable to match. Notably, the mass cost efficiency is demonstrated to be two orders of magnitude lower than that of classical control methods reported in the literature. These developments represent a significant advancement in active flow control in turbulent regimes, critical for industrial applications.

PMID:40406451 | PMC:PMC12092499 | DOI:10.1007/s10494-025-00642-x

Matern Fetal Med. 2022 Apr 26;4(2):103-112. doi: 10.1097/FM9.0000000000000147. eCollection 2022 Apr.

ABSTRACT

OBJECTIVE: This study investigates the efficacy of analyzing fetal heart rate (FHR) signals based on Artificial Intelligence to obtain a baseline calculation and identify accelerations/decelerations in the FHR through electronic fetal monitoring during labor.

METHODS: A total of 43,888 cardiotocograph(CTG) records of female patients in labor from January 2012 to December 2020 were collected from the NanFang Hospital of Southern Medical University. After filtering the data, 2341 FHR records were used for the study. The ObVue fetal monitoring system, manufactured by Lian-Med Technology Co. Ltd., was used to monitor the FHR signals for these pregnant women from the beginning of the first stage of labor to the end of delivery. Two obstetric experts together annotated the FHR signals in the system to determine the baseline as well as accelerations/decelerations of the FHR. Our cardiotocograph network (CTGNet) as well as traditional methods were then used to automatically analyze the baseline and acceleration/deceleration of the FHR signals. The results of calculations were compared with the annotations provided by the obstetric experts, and ten-fold cross-validation was applied to evaluate them. The root-mean-square difference (RMSD) between the baselines, acceleration F-measure (Acc.F-measure), deceleration F-measure (Dec.F-measure), and the morphological analysis discordance index (MADI) were used as evaluation metrics. The data were analyzed by using a paired t-test.

RESULTS: The proposed CTGNet was superior to the best traditional method, proposed by Mantel, in terms of the RMSD.BL (1.7935 ± 0.8099 vs. 2.0293 ± 0.9267, t = -3.55 , P = 0.004), Acc.F-measure (86.8562 ± 10.9422 vs. 72.2367 ± 14.2096, t = 12.43, P <0.001), Dec.F-measure (72.1038 ± 33.2592 vs. 58.5040 ± 38.0276, t = 4.10, P <0.001), SI (34.8277±20.9595 vs. 54.8049 ± 25.0265, t = -9.39, P <0.001), and MADI (3.1741 ± 1.9901 vs. 3.7289 ± 2.7253, t = -2.74, P = 0.012). The proposed CTGNet thus had significant advantages over the best traditional method on all evaluation metrics.

CONCLUSION: The proposed Artificial Intelligence-based method CTGNet delivers good performance in terms of the automatic analysis of FHR based on cardiotocograph data. It promises to be a key component of smart obstetrics systems of the future.

PMID:40406444 | PMC:PMC12094348 | DOI:10.1097/FM9.0000000000000147

Health Inf Sci Syst. 2025 May 21;13(1):37. doi: 10.1007/s13755-025-00354-6. eCollection 2025 Dec.

ABSTRACT

Early diagnosis and precise treatment of gastrointestinal (GI) diseases are crucial for reducing mortality and improving quality of life. In this context, the detection and classification of abnormalities in endoscopic images is an important support for specialists during the diagnostic process. In this study, an innovative deep learning approach for the segmentation and classification of pathological regions in the GI system is presented. In the first phase of the study, a novel segmentation network called GISegNet was developed. GISegNet is a deep learning-based architecture tailored for accurate detection of anomalies in the GI system. Experiments conducted on the Kvasir dataset showed that GISegNet achieved excellent results on performance metrics such as Jaccard and Dice coefficients and outperformed other segmentation models with a higher accuracy rate (93.16%). In the second phase, a hybrid deep learning method was proposed for classifying anomalies in the GI system. The features extracted from the transformer-based models were fused and optimized using the Minimum Redundancy Maximum Relevance (mRMR) algorithm. The classification process was performed using Support Vector Machines (SVM). As a result of feature fusion and selection, the second model, which combined features from DeiT and ViT models, achieved the best performance with an accuracy rate of 95.2%. By selecting a subset of 300 features optimized by the mRMR algorithm, the accuracy (95.3%) was maintained while optimizing the computational cost. These results show that the proposed deep learning approaches can serve as reliable tools for the detection and classification of diseases of the GI system.

PMID:40406365 | PMC:PMC12095780 | DOI:10.1007/s13755-025-00354-6

Matern Fetal Med. 2024 Jul 18;6(3):141-146. doi: 10.1097/FM9.0000000000000236. eCollection 2024 Jul.

ABSTRACT

OBJECTIVE: To determine whether deep learning algorithms are suitable for predicting preterm birth.

METHODS: A retrospective study was conducted at Peking University Third Hospital from January 2018 to June 2023. Birth data were divided into two parts based on the date of delivery: the first part was used for model training and validation, while real world viability was evaluated using the second part. Four machine learning algorithms (logistic regression, random forest, support vector machine, and transformer) were employed to predict preterm birth. Receiver operating characteristic curves were plotted, and the area under the curve (AUC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy were calculated.

RESULTS: This research included data on 30,965 births, where 24,770 comprised the first part, and included 3164 (12.77%) in the preterm birth group, with 6195 in the second part, including 795 (12.83%) in the preterm birth group. Significant differences in various factors were observed between the preterm and full-term birth groups. The transformer model (AUC = 79.20%, sensitivity = 73.67%, specificity = 72.48%, PPV = 28.21%, NPV = 94.95%, and accuracy = 72.61% in the test dataset) demonstrated superior performance relative to logistic regression (AUC = 77.96% in the test dataset), support vector machine (AUC = 71.70% in the test dataset), and random forest (AUC = 75.09% in the test dataset) approaches.

CONCLUSION: This study highlights the promise of deep learning algorithms, specifically the transformer algorithm, for predicting preterm birth.

PMID:40406277 | PMC:PMC12087897 | DOI:10.1097/FM9.0000000000000236

Anal Chem. 2025 May 22. doi: 10.1021/acs.analchem.4c06905. Online ahead of print.

ABSTRACT

Light scattering is a powerful tool that has been widely applied in various scenarios, such as nanoparticle analysis, single-cell measurement, and blood flow monitoring. However, noise is always a concerning and challenging issue in light scattering imaging (LSI) due to the complexity of noise sources. In this work, a deep learning-based adaptive denoising framework has been established to explore the temporal information on LSI videos, aiming to provide an unsupervised and self-learning denoising strategy for various application scenarios of LSI. This novel framework consists of three stages: noise distribution maps for describing the characteristics of LSI noise, video denoising based on the unsupervised learning of the FastDVDNet network, and denoising effect discrimination to screen the best denoised result for further processing. The denoising performance is validated by two common LSI applications: nanoparticle analysis and label-free identification of single cells. The result shows that our method compares favorably to existing methods in suppressing the background noise and enhancing the signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) of LSI. Consequently, the successful analysis of both particle size distribution and cell classification can be notably improved. The proposed unsupervised adaptive denoising method is expected to offer a powerful tool toward a fully automated denoising and improved accuracy in extensive applications of LSI.

PMID:40405330 | DOI:10.1021/acs.analchem.4c06905

J Cheminform. 2025 May 22;17(1):80. doi: 10.1186/s13321-025-01004-5.

ABSTRACT

The field of molecular representation has witnessed a shift towards models trained on molecular structures represented by strings or graphs, with chemical information encoded in nodes and bonds. Graph-based representations offer a more realistic depiction and support 3D geometry and conformer-based augmentation. Graph Neural Networks (GNNs) and Graph-based Transformer models (GTs) represent two paradigms in this field, with GT models emerging as a flexible alternative. In this study, we compare the performance of GT models against GNN models on three datasets. We explore the impact of training procedures, including context-enriched training through pretraining on quantum mechanical atomic-level properties and auxiliary task training. Our analysis focuses on sterimol parameters estimation, binding energy estimation, and generalization performance for transition metal complexes. We find that GT models with context-enriched training provide on par results compared to GNN models, with the added advantages of speed and flexibility. Our findings highlight the potential of GT models as a valid alternative for molecular representation learning tasks.

PMID:40405272 | DOI:10.1186/s13321-025-01004-5

Biomed Eng Online. 2025 May 22;24(1):64. doi: 10.1186/s12938-025-01396-3.

ABSTRACT

BACKGROUND: Neurological disorders, ranging from common conditions like Alzheimer's disease that is a progressive neurodegenerative disorder and remains the most common cause of dementia worldwide to rare disorders such as Angelman syndrome, impose a significant global health burden. Altered facial expressions are a common symptom across these disorders, potentially serving as a diagnostic indicator. Deep learning algorithms, especially convolutional neural networks (CNNs), have shown promise in detecting these facial expression changes, aiding in diagnosing and monitoring neurological conditions.

OBJECTIVES: This systematic review and meta-analysis aimed to evaluate the performance of deep learning algorithms in detecting facial expression changes for diagnosing neurological disorders.

METHODS: Following PRISMA2020 guidelines, we systematically searched PubMed, Scopus, and Web of Science for studies published up to August 2024. Data from 28 studies were extracted, and the quality was assessed using the JBI checklist. A meta-analysis was performed to calculate pooled accuracy estimates. Subgroup analyses were conducted based on neurological disorders, and heterogeneity was evaluated using the I2 statistic.

RESULTS: The meta-analysis included 24 studies from 2019 to 2024, with neurological conditions such as dementia, Bell's palsy, ALS, and Parkinson's disease assessed. The overall pooled accuracy was 89.25% (95% CI 88.75-89.73%). High accuracy was found for dementia (99%) and Bell's palsy (93.7%), while conditions such as ALS and stroke had lower accuracy (73.2%).

CONCLUSIONS: Deep learning models, particularly CNNs, show strong potential in detecting facial expression changes for neurological disorders. However, further work is needed to standardize data sets and improve model robustness for motor-related conditions.

PMID:40405223 | DOI:10.1186/s12938-025-01396-3

BMC Public Health. 2025 May 22;25(1):1884. doi: 10.1186/s12889-025-23059-7.

ABSTRACT

INTRODUCTION: Epidemic modeling is crucial for understanding and predicting infectious disease spread. To capture the complexity of real-world transmission, dynamic interactions between individuals with spatial heterogeneity must be considered. This modeling requires high-dimensional epidemic parameters, which can lead to unidentifiability; therefore, integrating various data types for inference is essential to effectively address these challenges.

METHODS: We introduce a novel hybrid framework, Multi-Patch Model Update with Graph Attention Network (MPUGAT), that combines a multi-patch compartmental model with a spatio-temporal deep learning model. MPUGAT employs a GAT (Graph Attention Mechanism) to transform static traffic matrices into dynamic transmission matrices by analyzing patterns in diverse time series data from each city.

RESULTS: We demonstrate the effectiveness of MPUGAT through its application to COVID-19 data from South Korea. By accurately estimating time-varying transmission rates, MPUGAT outperforms traditional models and aligns with actual policies such as social distancing.

CONCLUSION: MPUGAT offers a novel approach for effectively integrating easily accessible, low-dimensional, non-epidemic-related data into epidemic modeling frameworks. Our findings highlight the importance of incorporating dynamic data and utilizing graph attention mechanisms to enhance accuracy of infectious disease modeling and the analysis of policy interventions. This study underscores the potential of leveraging diverse data sources and advanced deep learning techniques to improve epidemic forecasting and inform public health strategies.

PMID:40405112 | DOI:10.1186/s12889-025-23059-7

Eur Radiol. 2025 May 22. doi: 10.1007/s00330-025-11696-w. Online ahead of print.

ABSTRACT

OBJECTIVES: To evaluate the variability of fully automated airway quantitative CT (QCT) measures caused by different kernels and the effect of kernel conversion.

MATERIALS AND METHODS: This retrospective study included 96 patients who underwent non-enhanced chest CT at two centers. CT scans were reconstructed using four kernels (medium soft, medium sharp, sharp, very sharp) from three vendors. Kernel conversion targeting the medium soft kernel as reference was applied to sharp kernel images. Fully automated airway quantification was performed before and after conversion. The effects of kernel type and conversion on airway quantification were evaluated using analysis of variance, paired t-tests, and concordance correlation coefficient (CCC).

RESULTS: Airway QCT measures (e.g., Pi10, wall thickness, wall area percentage, lumen diameter) decreased with sharper kernels (all, p < 0.001), with varying degrees of variability across variables and vendors. Kernel conversion substantially reduced variability between medium soft and sharp kernel images for vendors A (pooled CCC: 0.59 vs. 0.92) and B (0.40 vs. 0.91) and lung-dedicated sharp kernels of vendor C (0.26 vs. 0.71). However, it was ineffective for non-lung-dedicated sharp kernels of vendor C (0.81 vs. 0.43) and showed limited improvement in variability of QCT measures at the subsegmental level. Consistent airway segmentation and identical anatomic labeling improved subsegmental airway variability in theoretical tests.

CONCLUSION: Deep learning-based kernel conversion reduced the measurement variability of airway QCT across various kernels and vendors but was less effective for non-lung-dedicated kernels and subsegmental airways. Consistent airway segmentation and precise anatomic labeling can further enhance reproducibility for reliable automated quantification.

KEY POINTS: Question How do different CT reconstruction kernels affect the measurement variability of automated airway measurements, and can deep learning-based kernel conversion reduce this variability? Findings Kernel conversion improved measurement consistency across vendors for lung-dedicated kernels, but showed limited effectiveness for non-lung-dedicated kernels and subsegmental airways. Clinical relevance Understanding kernel-related variability in airway quantification and mitigating it through deep learning enables standardized analysis, but further refinements are needed for robust airway segmentation, particularly for improving measurement variability in subsegmental airways and specific kernels.

PMID:40405045 | DOI:10.1007/s00330-025-11696-w

Int J Comput Assist Radiol Surg. 2025 May 22. doi: 10.1007/s11548-025-03406-0. Online ahead of print.

ABSTRACT

PURPOSE: Assessment of intraoperative surgical skill is necessary to train surgeons and certify them for practice. The generalizability of deep learning models for video-based assessment (VBA) of surgical skill has not yet been evaluated. In this work, we evaluated one unsupervised domain adaptation (UDA) and three semi-supervised (SSDA) methods for generalizability of models for VBA of surgical skill in capsulorhexis by training on one dataset and testing on another.

METHODS: We used two datasets, D99 and Cataract-101 (publicly available), and two state-of-the-art models for capsulorhexis. The models include a convolutional neural network (CNN) to extract features from video images, followed by a long short-term memory (LSTM) network or a transformer. We augmented the CNN and the LSTM with attention modules. We estimated accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC).

RESULTS: Maximum mean discrepancy (MMD) did not improve generalizability of CNN-LSTM but slightly improved CNN transformer. Among the SSDA methods, Group Distributionally Robust Supervised Learning improved generalizability in most cases.

CONCLUSION: Model performance improved with the domain adaptation methods we evaluated, but it fell short of within-dataset performance. Our results provide benchmarks on a public dataset for others to compare their methods.

PMID:40405033 | DOI:10.1007/s11548-025-03406-0

Sci Rep. 2025 May 22;15(1):17875. doi: 10.1038/s41598-025-02889-w.

ABSTRACT

Sepsis remains a leading cause of mortality in critical care settings, necessitating timely and accurate risk stratification. However, existing machine learning models for sepsis prediction often suffer from poor interpretability, limited generalizability across diverse patient populations, and challenges in handling class imbalance and high-dimensional clinical data. To address these gaps, this study proposes a novel framework that integrates bio-inspired feature selection and graph-based deep learning for enhanced sepsis risk prediction. Using the MIMIC-IV dataset, we employ the Wolverine Optimization Algorithm (WoOA) to select clinically relevant features, followed by a Generative Pre-Training Graph Neural Network (GPT-GNN) that models complex patient relationships through self-supervised learning. To further improve predictive accuracy, the TOTO metaheuristic algorithm is applied for model fine-tuning. SMOTE is used to balance the dataset and mitigate bias toward the majority class. Experimental results show that our model outperforms traditional classifiers such as SVM, XGBoost, and LightGBM in terms of accuracy, AUC, and F1-score, while also providing interpretable mortality indicators. This research contributes a scalable and high-performing decision support tool for sepsis risk stratification in real-world clinical environments.

PMID:40404796 | DOI:10.1038/s41598-025-02889-w