Computers in biology and medicine最新文献

Background and motivation: Classifying diseases like heart problems using gene expression data depends on selecting important genes. Traditional machine learning (ML) often uses simple feature selection (FS) techniques, which can limit accuracy. In our research, we combine deep learning (DL) with gene-focused methods like differential expression analysis (DEA) to improve classification performance significantly.

Method: We thoroughly and rigorously evaluated ML and DL classifiers using two gene expression datasets (GSE36961 and GSE57345). We tested four hypotheses using feature selection methods such as chi-square, DEA. We applied principal component analysis (PCA) to reduce the number of features. To ensure the reliability of our findings, we applied k-fold cross-validation, hyperparameter tuning, block effect analysis, and assessed data augmentation and generalization. Statistical tests, including paired t-test and Mann-Whitney U test, and Wilcoxon signed-rank test were performed to compare model performances rigorously.

Results: Our experiments on two gene expression datasets (GSE36961, GSE57345) not only confirmed all four hypotheses (H1, H2, H3, and H4) but also revealed significant performance improvements. For H1, without FS, DL outperformed ML models by a substantial margin. For H2, with FS, DL outperformed ML models by a significant percentage. In H3, ML with FS improved over ML without FS by a considerable margin. For H4, DL with FS outperformed DL without FS by a noticeable percentage. Among FS methods, DEA consistently yielded the best results for both ML and DL, further underlining the significance of our findings.

Conclusions: Combining DL with biological feature selection, especially DEA, improves gene expression classification and enables gene ranking and biomarker identification. This integrative approach balances modeling power with biological relevance, providing a reproducible framework for robust biomarker-based classification.

As they navigate complex extracellular environments, cells and their nuclei undergo extensive deformation. Recent experiments have demonstrated that vascular endothelial cells cultured on microgroove substrates, which mimic the anisotropic topography of the basement membrane, exhibit complex nuclear deformations, leading to partial or even complete nuclear penetration into the microgrooves. Interestingly, the experiments suggest that nuclear entry into the microgrooves is driven mainly by cellular adhesion and spreading rather than by cytoskeleton-mediated pulling and/or pushing forces. In the present work, we develop a phase-field model to describe endothelial cell deformation on microgroove substrates and characterize the conditions necessary for nuclear confinement within the grooves, a process that has been termed "caging" in the experiments. The model introduces a novel non-local term that prevents the cellular body from fragmenting under conditions of strong adhesion and high curvature. Our numerical simulations show that significant nuclear deformation and partial caging occur for strong cell-substrate adhesion and for nuclear membrane stiffness close to or inferior to that of the cell membrane. We further show that the dimensions of the grooves are critical for the caging process, with increasing groove depth and width favoring nuclear penetration into and caging within the grooves. These results are in close agreement with experimental observations, thus corroborating the notion that cell-substrate adhesion forces can drive large-scale nuclear deformations without the need for cytoskeleton-generated forces.

Objectives: This study proposes a deep learning framework and an annotation methodology for the automatic detection of periodontal bone loss landmarks, associated conditions, and staging. Methods192 periapical radiographs were collected and annotated using a stage agnostic methodology, labelling clinically relevant landmarks regardless of disease presence or extent. We propose a heuristic post-processing module that aligns predicted keypoints to tooth boundaries using an auxiliary instance segmentation model. An evaluation metric, Percentage of Relative Correct Keypoints (PRCK), is proposed to capture keypoint performance in dental imaging domains. Four donor pose estimation models were adapted with fine-tuning for our keypoint problem. Results Post-processing improved fine-grained localisation, raising average PRCK^0.05 by +0.028, but reduced coarse performance for PRCK^0.25 by -0.0523 and PRCK^0.5 by -0.0345. Orientation estimation shows excellent performance for auxiliary segmentation when filtered with either stage 1 object detection model. Periodontal staging was detected sufficiently, with the best mesial and distal Dice scores of 0.508 and 0.489, while furcation involvement and widened periodontal ligament space tasks remained challenging due to scarce positive samples. Scalability is implied with similar validation and external set performance. Conclusion The annotation methodology enables stage agnostic training with balanced representation across disease severities for some detection tasks. The PRCK metric provides a domain-specific alternative to generic pose metrics, while the heuristic post-processing module consistently corrected implausible predictions with occasional catastrophic failures.

Clinical significance: The proposed framework demonstrates the feasibility of clinically interpretable periodontal bone loss assessment, with the potential to reduce diagnostic variability and clinician workload.

Objective: Detecting Alzheimer's disease (AD) at an early stage is essential for administering effective treatments and preventing neuronal damage. Unfortunately, current diagnostic techniques are often invasive and expensive. Our research focuses on creating a cost-effective and non-invasive method for the early detection of cognitive decline.

Methods: Using a publicly available dataset of resting state electroencephalographic (EEG) data on healthy controls and patients with Mild Cognitive Impairment (MCI), two novel deep learning (DL) algorithms with self-attention mechanisms were developed and evaluated for their performance in predicting MCI and cognitive decline.

Results: Both proposed DL algorithms outperformed a traditional convolutional neural network (CNN) model in predicting MCI, achieving test accuracy improvements of 8.5% and 10%, respectively, while utilizing significantly fewer trainable parameters. An ablation study highlighted the attention layer as a key feature, enhancing model accuracy by 8.5%. Analysis of the attention layers indicated that beta band frequencies (13-30 Hz) were essential for distinguishing MCI from control subjects, highlighting the role of high EEG frequencies in early cognitive deficits. Predicting pre-clinical cognitive decline in healthy subjects proved more challenging than predicting diagnosed MCI. However, using transfer-learning methods, we achieved a test accuracy of 56.08%.

Conclusion: Our models achieved state-of-the-art results in the MCI classification task, and demonstrated learning progress in predicting cognitive decline in the preclinical stage. As this is the first time DL models have been evaluated to classify healthy subjects based on cognitive scores, where brain changes are minimal and difficult to detect, this study opens new avenues for discovering biomarkers in early AD diagnosis and facilitating early interventions. Interpretation of the trained DL attention models provided valuable insights that aligned with the existing brain research, serving as a helpful tool for validating AI in healthcare applications.

Intracerebroventricular (ICV) streptozotocin (STZ) deveops Alzheimer's disease (AD)-like conditions in rodents, which are characterized by insulin resistance, tau pathology, and neurodegeneration. Hentriacontane, a natural compound found in various sources, including beeswax, possesses anti-inflammatory and antioxidant properties. In the present investigation, we performed in silico molecular docking, molecular dynamics, MMGBSA, PCA, and FEL analysis of hentriacontane and rivastigmine with acetylcholinesterase (AchE). Further, we assessed the in vivo neuroprotective effects of hentriacontane in an ICV-STZ-induced AD-like condition in rats. STZ (3 mg/kg/ICV) was injected into male Sprague-Dawley rats. Cognitive functions were evaluated by Barnes-Maze (BM), novel object recognition test (NORT), and passive avoidance test (PAT). Hentriacontane (3 and 5 mg/kg) and rivastigmine (1 mg/kg) were given intraperitoneally for 14 days. Brain-derived neurotrophic factor (BDNF), AchE, oxidative stress parameters including GSH, MDA, SOD, and CAT, and proinflammatory cytokines including IL-6, TNF-α, IL-1β, and NF-ҡB were measured via ELISA. Further, we have also estimated the BACE1 and NO levels. Histopathological evaluation was conducted using hematoxylin and eosin staining. In silico molecular docking, dynamics, and post-dynamics data revealed promising binding affinities of hentriacontane for AchE. Further, hentriacontane attenuated ICV-STZ-induced cognitive deficit in BM, NORT, and PAT. Additionally, altered oxidative stress, proinflammatory, and cell signalling parameters were restored. Histopathology revealed that the hentriacontane-treated group showed significant restoration of the small pyramidal cells in the CA1 and CA2 regions of the brain. Hentriacontane demonstrated neuroprotective effects by modulation of AchE, leading to improved cognitive functions as evidenced by in silico and in vivo investigations.

Background: Pancreatic ductal adenocarcinoma (PDAC) is a highly lethal malignancy, with accurate preoperative assessment of vascular involvement critical for determining resectability and treatment planning. Conventional contrast-enhanced CT relies on qualitative evaluations, leading to interobserver variability and diagnostic uncertainty. Existing radiomics studies for PDAC mostly focus on single anatomical structures and lack organ-level interpretability, limiting clinical translation.

Methods: A retrospective study was conducted using the international PANORAMA CT cohort, with 1488 eligible samples stratified into PDAC diagnosis (1186 cases) and vascular involvement prediction (302 cases) tasks. Standardized radiomic features were extracted from five key structures (artery, vein, pancreatic parenchyma, pancreatic duct, common bile duct) following IBSI guidelines. After LASSO-based dimensionality reduction, six machine learning classifiers were trained for each structure, with top-performing models integrated into structure-specific consensus models. A meta-level consensus model was constructed via stacking, and SHAP analysis was applied for organ-level interpretability. Model performance was evaluated using AUC, accuracy, calibration curves, and decision curve analysis (DCA).

Results: The multi-structure consensus model achieved an AUC of 0.975 (95% CI: 0.956-0.990) with 0.937 accuracy for PDAC diagnosis, and an AUC of 0.868 (95% CI: 0.769-0.952) with 0.803 accuracy for vascular involvement prediction in independent testing cohorts. DeLong tests demonstrated the model significantly outperformed four single-structure models (artery, vein, pancreatic duct, common bile duct) in both tasks (all P < 0.05), with no significant difference compared to the pancreas parenchyma model (PDAC diagnosis: P = 0.078; vascular involvement prediction: P = 0.093). SHAP analysis identified pancreatic parenchyma as the dominant contributor to PDAC diagnosis and arterial features as key for vascular involvement prediction. The model exhibited robust calibration (MAE = 0.01 for PDAC; MAE = 0.02 for vascular involvement) and clinical net benefit via DCA.

Conclusion: The proposed multi-structure CT radiomics consensus model integrates contextual information from multiple pancreatic structures, achieving competitive performance for PDAC diagnosis and vascular involvement prediction. Organ-level SHAP interpretation enhances clinical transparency, offering a reliable tool to support preoperative decision-making in PDAC.