Computer methods and programs in biomedicine最新文献

Background & objective: Most existing methods for indirectly deriving reference intervals (RIs) from routine laboratory databases use univariate approaches with limited or no rigorous data cleaning. Recognizing the potential of multivariate data-mining strategies, we developed novel software-SOM-clean-that employs self-organizing map (SOM) clustering for iterative exclusion of records exhibiting atypical multi-test patterns.

Methods: We retrieved records for 22 major health-screening tests (HSTs) from a Saudi Arabian laboratory participating in a RI study. After excluding records from frequently tested individuals and those with <10 HST results, 37,285 records remained for analysis. Initial crude RIs were calculated parametrically using a two-parameter Box-Cox power transformation. All transformed values were standardized against these RIs to generate uniform-scale values, so that any result within RI limits fell between ±1.96. The self-organizing map (m × m cells, m = 5-8) was initialized with normal random values, and records were clustered into cells with highest similarity. Cells' patterns were updated by records assigned to each of them. This learning process of the map was repeated until equilibrium. Subsequently, cells exhibiting atypical features were excluded, and RIs were recalculated using records from the remaining cells. This process was repeated iteratively until all RIs stabilized.

Results: Histograms of retrieved results frequently exhibited peaks differing in shape and location from those in the direct study (n = 880). The goodness-of-fit (GOF) of SOM-clean RIs was assessed by skewness, kurtosis, and Kolmogorov-Smirnov test P-values after transformation, as well as by the bias ratio of reference limits compared with the direct study. GOF depended on map size and criteria for identifying atypical cells; the software therefore incorporated an all-inclusive search for optimal conditions referencing the direct study RIs. By using the optimal settings, SOM-clean achieved excellent GOF of RIs simultaneously across nearly all HSTs, indicating conformity of the estimated RIs to the healthy status. In comparison, RIs derived using a representative indirect method (refineR) were generally broader or biased, particularly for tests with highly skewed distributions.

Conclusion: SOM-clean represents a practical and robust parametric tool for estimating RIs indirectly from routine laboratory data employing a novel multivariate-based data cleaning scheme.

Background and objective: In Brazil, coronary angioplasty with stent implantation is a primary intervention for cardiovascular diseases, yet in-stent restenosis remains a significant complication. Recent proposals suggest transitioning from traditional cylindrical stents to conical geometries to better align with vascular physiology. This study aims to compare the performance of cylindrical and conical stents and investigate the influence of varying strut thicknesses on hemodynamic parameters.

Methods: The study employed computational modeling using both Fluid-Structure Interaction (FSI) and Computational Fluid Dynamics (CFD) simulations to quantify hemodynamic parameters including Time-Averaged Wall Shear Stress (TAWSS), Oscillatory Shear Index (OSI), and Relative Residence Time (RRT). A total of 12 simulations were performed (6 FSI and 6 CFD) on models of cylindrical and conical arteries with stent strut thicknesses ranging from 0.1 mm to 0.3 mm. The finite volume method was used for the fluid domain, while the finite element method was applied to the solid domain (arterial wall and stent). Blood was modeled as a non-Newtonian fluid using the Carreau model, with Reynolds numbers from 251 to 381 and Womersley numbers from 2.23 to 3.78.

Results: Quantitative analysis revealed that rigid-wall CFD consistently underestimates the risk of restenosis compared to FSI. Specifically, FSI predicted areas of critical Time-Averaged Wall Shear Stress (TAWSS ≤ 1 Pa) that were 12% to 46% larger than those predicted by CFD. Strut thickness emerged as a dominant factor; increasing thickness to 0.3 mm resulted in WSS values approximately three times lower than the 0.1 mm models, significantly expanding recirculation zones. Regarding geometry, while cylindrical stents exhibited concentrated high Oscillatory Shear Index (OSI) at the distal edge, conical stents demonstrated a more distributed OSI pattern and a markedly improved Relative Residence Time profile, reducing peak RRT at the distal edge by approximately 60% compared to cylindrical models (12.25Pa^-1 vs. 29.94Pa^-1), thereby mitigating stagnation and potential edge restenosis.

Conclusions: The findings confirm that neglecting arterial compliance (CFD only) leads to a substantial underestimation of hemodynamic risk. Both stent geometry and strut thickness are critical; while conical stents offer better risk distribution, thicker struts can negate these benefits. Optimizing these parameters is essential for next-generation stent designs.

Background and objective: Low birth weight (LBW) is a major global public health concern, strongly linked to neonatal morbidity and long-term health complications. Early prediction of LBW is essential to reduce neonatal mortality and guide targeted healthcare interventions. This study proposes a predictive framework integrating machine learning (ML), deep learning (DL), and model-agnostic eXplainable Artificial Intelligence (XAI) to identify key socioeconomic and demographic determinants of LBW.

Methods: Data from the Bangladesh Demographic and Health Survey (BDHS), comprising 1574 participants and 12 variables, are analyzed. Key predictors included maternal age, education, household wealth, geographic region, birth order, and maternal BMI. Chi-square tests assess variable associations. A stacking ensemble model, SmartFusion-LR5, is developed, combining K-Nearest Neighbors, Logistic Regression (LR), Decision Tree, Random Forest, and Naive Bayes, with LR as the meta-learner. Model performance is evaluated using accuracy, precision, recall, area under the curve (AUC), F1-score, and Matthews correlation coefficient (MCC).

Results: Significant disparities in LBW prevalence are observed across geographic divisions, with higher parental education and socioeconomic status associated with healthier outcomes. The SmartFusion-LR5 model achieves the highest overall discriminative capability compared to baselines, attaining 93.0% accuracy, 86.7% precision, 99.8% recall, 92.8% F1-score, 94.0% AUC, and an MCC of 86.0%. Comparable performance also obtained from SmartFusion-XGB4 (91.8% accuracy, 86.2% precision, 99.6% recall, 92.4% F1-score, 92.2% AUC, MCC 84.8%) and SmartFusion-RF4 (91.7% accuracy, 86.2% precision, 99.2% recall, 92.2% F1-score, 92.0% AUC, MCC 84.4%). Global XAI methods identified age at first birth, division, residence, wealth, and husband's education as key determinants, while local explanations revealed individual feature impacts.

Conclusions: The proposed framework offers a robust, interpretable, and scalable approach for early LBW risk prediction, supporting targeted maternal and child health interventions in resource-constrained settings.

Background: Intracerebral hemorrhage (ICH) is a leading cause of long-term disability, particularly in China. Post-stroke motor recovery exhibits considerable heterogeneity, presenting substantial challenges for clinicians in establishing realistic goals. While integrative AI paradigms have shown success in other complex medical domains, existing models often rely on single-modality data, limiting their predictive accuracy and clinical utility. This study aims to develop and validate a multimodal predictive model integrating CT imaging, clinical data, and rehabilitation assessments to simultaneously predict motor recovery and global rehabilitation outcomes following cerebral hemorrhage.

Methods: We conducted a retrospective study involving 739 patients (315 for motor function prediction and 424 for rehabilitation outcome assessment) who received rehabilitation therapy after cerebral hemorrhage. To predict motor function, we constructed a late-fusion deep learning model leveraging 3D-DenseNet for CT neuroimaging and Multi-Layer Perceptron (MLP) for clinical and laboratory features. To predict rehabilitation outcomes, a Gradient Boosting Decision Tree (GBDT) model was developed and validated using 5-fold and 10-fold cross-validation, comparing it against other machine learning algorithms, including SVR, Random Forest and AdaBoost. Model performance was assessed using metrics including AUC and R². Additionally, univariate and multivariate regression analysis were performed to identify significant factors influencing motor recovery and rehabilitation outcomes.

Results: A total of 739 patients were included. The multimodal fusion model achieved an AUC of 0.856 (95 % CI: 0.741-0.971) and an F1 score of 0.897 (95 % CI: 0.819-0.975), significantly outperforming the imaging-only (AUC: 0.833) and clinical-only (AUC: 0.749) models. For rehabilitation outcome prediction, the GBDT model achieved an R² of 0.849 (95 % CI: 0.803-0.887), demonstrating superior stability and accuracy over other models. Additionally, multivariate analysis revealed that serum albumin (ALB), neutrophil percentage (NEUT%), triglycerides (TG), and thrombin time (TT) were independent predictors of motor recovery, while age, admission mBI, and time to start rehabilitation significantly influenced functional outcomes.

Conclusion: This study confirms that a multimodal deep learning framework integrating routinely available CT imaging and clinical biomarkers provides high predictive value for simultaneously forecasting motor recovery and global functional outcomes after ICH. This proof-of-concept approach offers a reproducible, data-driven tool for early risk stratification, facilitating the formulation of individualized rehabilitation strategies and optimizing resource allocation in clinical workflows.

Background and objective: Venoarterial extracorporeal membrane oxygenation (VA ECMO) circuits typically utilise a continuous flow (CF) of blood to support patients suffering from refractory cardiorespiratory dysfunction. Pulsatile flow (PF) VA ECMO is an emerging technology being developed to overcome adverse effects associated with non-physiological CF VA ECMO such as worsening of microcirculatory and cardiac function. However, the flow dynamics associated with PF VA ECMO, such as positioning of the watershed region, wall shear stress, and ventricular unloading are still largely unknown. Therefore, to address this gap, our study aimed to utilise computational fluid dynamics (CFD) to compare the arterial cannula flow characteristics generated by CF and PF VA ECMO.

Methods: A multiscale CFD model was created using a patient-specific aortic geometry and employed a closed-loop lumped parameter network as boundary conditions. Mean VA ECMO flow rates of 3, 4, and 5 L/min were simulated for both CF and counter-pulsed PF scenarios.

Results: The hemodynamic results demonstrated increased stroke volume, ejection fraction, and coronary flow during PF VA ECMO, and decreased left ventricular volumes, afterload, and pressure-volume areas, when compared to CF VA ECMO. Delivery of oxygen saturated blood from VA ECMO to the upper body decreased slightly during PF VA ECMO during 4 L/min of support. Lastly, wall shear stress on the aortic wall increased substantially during PF VA ECMO, when compared to CF VA ECMO.

Conclusions: The findings from this study suggest varied hemodynamic and flow dynamic outcomes when comparing CF and PF VA ECMO, each with their own benefits and drawbacks.

Background and objective: Cardiac magnetic resonance imaging provides detailed anatomical information but is costly and not feasible for routine monitoring. Accurate control of cardiac substructure areas is essential for studying development, adaptation, and disease progression. This work introduces a framework for synthetic cardiac imaging that enables parameter-driven area modifications using oriented bounding box representations while preserving anatomical plausibility.

Methods: We propose a three-stage framework that integrates: (1) Encoding each cardiac substructure as an oriented bounding box to enable structured representation and easier shape manipulation. (2) A progressive label modification algorithm to apply parameter-driven area changes while maintaining anatomical consistency. (3) A bounding-box-to-segmentation model for reconstructing detailed masks and (4) A diffusion-based segmentation-to-image synthesis model for generating realistic cardiac magnetic resonance images. The oriented bounding box encoding serves as the foundation for controlled anatomical transformations, while the subsequent models ensure structural plausibility and image fidelity.

Results: Experiments show that oriented bounding box encoding enables more accurate control of cardiac substructure area modifications than conventional approaches. Both increments and decrements exhibit a systematic deviation of about 5%, which is effectively corrected by applying a 5% calibrated input, reducing mean errors to within ±1%. Generated images remain anatomically plausible and structurally consistent.

Conclusions: The proposed framework enables parameter-driven cardiac MRI synthesis with precise substructure area control. By combining oriented bounding box encoding, progressive modification, and diffusion modeling, it achieves anatomically consistent results while reducing reliance on repeated scans, supporting applications in longitudinal monitoring and progression studies.

Background and objective: Developing multimodal data-driven diagnostic systems has become a key clinical strategy for improving breast cancer outcomes. However, effectively modeling multimodal features remains challenging due to substantial semantic heterogeneity, scale discrepancies, and the inherent difficulty of cross-modal alignment. Although existing studies have proposed various multimodal fusion methods, most rely on direct feature concatenation or shallow integration, which fail to capture fine-grained intra-modality semantics as well as the complex interactions between histopathological and genomic modalities.

Methods: In this study, we propose a multimodal diagnostic framework based on Feature Enhancement and Semantic Collaborative Alignment (FESCA). The method incorporates a semantic-guided modality feature enhancement mechanism that effectively extracts and strengthens diagnostic cues from both pathological images and genomic data. In addition, a contrastive-learning-based cross-modal alignment strategy is introduced to map heterogeneous modalities into a unified semantic space and achieve deep semantic collaboration through contrastive optimization. To ensure robust breast cancer classification under varying modality availability, a multimodal collaborative diagnostic strategy is employed to dynamically adapt the feature representations.

Results: We evaluate FESCA on the TCGA-BRCA dataset, and the experimental results demonstrate that it outperforms state-of-the-art methods in breast cancer classification while significantly improving both intra-modality representation quality and cross-modal semantic alignment.

Conclusion: To enhance accessibility and practical application, we developed a web-based breast cancer pathological staging diagnosis system to visualize and deploy the FESCA model, demonstrating a step toward clinical application and providing a benchmark for other research methods.

Background: Endovascular aneurysm repair (EVAR) is preferred for abdominal aortic aneurysms (AAAs) due to minimal invasiveness, but complications like Type I endoleaks (T1EL) and intra-prosthetic thrombus (IPT) persist. While current risk prediction routinely involves anatomical assessment, the specific postoperative hemodynamic environments associated with these distinct complications remain under-investigated.

Methods: This study integrates patient-specific computed tomography (CT) angiography with a computational fluid dynamics (CFD)-based thrombus growth model to analyze postoperative geometries in 12 patients divided into three groups: T1EL, IPT, and no complications (Control). Key hemodynamic descriptors included flow patterns, wall shear stress (WSS), helicity, pressure gradients, and predicted thrombus distribution.

Results: T1EL patients showed complex flow patterns and distinct pressure gradients near the proximal or distal sealing zones. The IPT group exhibited significantly low time-averaged wall shear stress (TAWSS) and high relative residence time (RRT) within the graft limbs, fostering platelet aggregation. CFD simulations indicated different spatial distributions of potential thrombus formation between groups. Specifically, T1EL cases demonstrated stronger 3D helical flow features, while IPT cases had weaker rotational flow. Notably, graft geometry analysis revealed that cross-limb configurations induced specific flow disturbances distinct from standard configurations.

Conclusions: By combining patient-specific biomechanics and thrombus modeling, this study offers new insights into the distinct flow phenotypes of post-EVAR complications. It highlights hemodynamic descriptors-helicity, TAWSS, oscillatory shear index (OSI), RRT, and endothelial cell activation potential (ECAP)-as potential quantitative metrics to complement anatomical assessment. Findings also suggest cross-limb graft configurations may modulate flow dynamics, potentially increasing thrombotic risk, providing preliminary clinical guidance requiring validation.

Background and objective: The difference in molecular characteristics of Triple negative breast cancer (TNBC) aids in distinguishing between its four prominent subtypes- basal-like 1, basal-like 2, mesenchymal, and luminal androgen receptor. This study presents the first integrative framework that combines explainable AI with machine learning approaches to classify TNBC subtypes. Unlike conventional models, our approach offers interpretability while enabling biomarker prioritization by identifying key hub genes that drive subtype-specific predictions.

Methods: In the experiment 783 cases (BL1 (160), BL2 (75), M (151), LAR (106), non-TNBC (291) reported in Gene Expression Omnibus (GEO) and Genomic Data Commons (GDC) data portal were used for the analysis. The proposed framework comprises modules for the identification of gene signatures for the four-subtype followed by the classification model based on eight different machine learning algorithms. Random Forest classifier was found to be best model with 96 % testing accuracy, which was elected for Explainable framework using Shapley Additive Explanations.

Results: Explainable biomarker module could provide a set of 47 biomarkers which is relevant in distinguishing the four types on triple negative breast cancer. The biomarkers could have the potential to be considered for TNBC prognosis in clinical setting.

Conclusion: Key findings highlight the hub genes CDC20, CDCA2, PIMREG, KIF2C, and CENPW, implicating pathways such as ubiquitin-proteasome signaling and microtubule dynamics. These insights pave the way for biomarker-driven therapies and precision medicine in triple negative breast cancer.