Computers in Biology and Medicine

CVDs are heterogeneous, multifactorial disorders that remain the leading cause of global mortality from infancy to old age. It requires an early identification and treatment of risk factors to accelerate disease prevention and morbidity improvement. Advancements in transcriptomics technologies gives large pool of heterogenous gene expression data. The technical heterogeneity of gene expression data reduces ability to compare multiple cross-platform datasets at once. To bridge gap, we systematically evaluate three data harmonization techniques: Shambhala-2, TDM, and UPC to align heterogeneous data into a shared expression space while preserving biological signals. Our pipeline integrates 25 independent datasets comprising 983 samples across 23 distinct CVDs phenotypes from both RNA-seq and microarray platforms. The framework benchmarks 35 Machine learning (ML) and Deep learning (DL) classifiers, including Transformers and ResNets, across three data modalities such as RNA-seq, microarray hybridization and RNA-seq + microarray and multiple tissue types. To ensure clinical trustworthiness, we apply multiple Explainable artificial intelligence (XAI) methods, such as SHapley additive exPlanations (SHAP) and Integrated gradientss (IGs), and assess their reliability using quantitative metrics like Area over the perturbation curve (AOPC), Sensitivity, and Infidelity. Results indicate that Shambhala-2 provides superior harmonization by maximizing the biological signal-to-platform ratio. Evaluation of XAI methods reveals that Shapley-based approaches offer the highest stability for identifying influential genomic features in high-dimensional data. Functional enrichment and pathway analyses further confirmed the involvement of identified biomarkers in key cardiovascular processes, including inflammation, immune regulation, oxidative stress, and vascular remodeling. Collectively, this study provides a scalable and interpretable road-map that integrates XAI with cross-dataset biomarker discovery, supporting the transition toward precision cardiology.

Vector flow mapping (VFM) reconstructs left-ventricular (LV) blood velocity from color-Doppler echocardiography by combining the measured beamwise component with physical and regularizing constraints. Analysis of the discrete VFM formulation shows that the inverse problem is intrinsically ill posed: the occurrence of singular modes can be predicted from the geometry of the segmented blood-pool domain, the imposed boundary conditions, and the degree of smoothing. These modes can propagate uncertainty along entire transverse bands of the reconstructed velocity field, yet conventional VFM neither quantifies this uncertainty nor allows for correcting the blood-pool segmentation. We introduce Bayesian VFM (B--VFM), a hierarchical framework that jointly infers radial and transverse velocities, a probabilistic blood-pool mask, their spatially resolved uncertainties, and hyperparameters weighting Doppler and segmentation fidelity, mass conservation, boundary conditions, and smoothness. The discretized posterior admits a closed-form gradient and exact Hessian, enabling computationally efficient, gradient-based MAP estimation, sampling, and direct analysis of ill-posed modes. Posterior inference combines Gibbs sampling of conjugate Gamma-distributed hyperparameters with conditional maximum-a-posteriori estimation and a Laplace approximation for the high-dimensional velocity and mask fields. To accommodate systematic departures from planar mass conservation, B-VFM can learn the covariance of the planar divergence residual from an ensemble of flows and incorporate it as a structured model-discrepancy prior. Independent chains converged reproducibly, while covariance priors learned from flow ensembles illustrated how model discrepancies can be incorporated into the inference. B--VFM was evaluated using Lamb-Chaplygin dipoles under ideal conditions and with Doppler corruption, Doppler voids, and segmentation defects, and using the Hicks-Moffatt family of spherical vortices to assess violations of planar mass conservation. The method produced smooth reconstructions, localized uncertainty near unreliable measurements and regions of model inconsistency, and used flow information to correct segmentation errors. Within the tested vortex family, the data-informed planar divergence prior reduced velocity bias and mask distortion. B--VFM thus provides an uncertainty-aware reconstruction method and a flexible foundation for future VFM formulations incorporating additional priors, observations, and physical models. Future work will evaluate the method using clinical data and more complex three-dimensional benchmark flows.

Age and sex estimation from dental radiographs supports forensic identification and age-dependent legal assessment. Multi-input models may combine global and regional representations, but when these inputs are derived from the same radiograph, they may repeat anatomical information rather than provide complementary signals. We developed HMA-Net, a hierarchical multi-stream anatomical network, together with a task-specific input audit that evaluates whether each candidate representation improves at least one prediction task without materially degrading the other. The model was developed using a retrospective single-centre cohort of 4,293 panoramic radiographs from individuals aged 3-18 years and combined a panoramic-image stream with graph-linked tooth representations. On the held-out test set, the final panoramic-plus-permanent-tooth model achieved a mean absolute age error of 0.51 years and a sex-classification accuracy of 85.2%; 99.4% of age estimates were within 2 years of chronological age. Across three independent training runs, removing the jaw-region stream reduced mean absolute age error from 0.68 +/- 0.01 to 0.52 +/- 0.01 years and increased sex accuracy from 81.3+/- 1.7% to 85.8 +/- 0.8%. Adding 20 deciduous-tooth classes provided no age benefit and reduced sex accuracy by 7.9% relative to the permanent-only representation. The mandibular first molars received the greatest tooth-level attention, and their occlusion impaired both prediction tasks. These findings show that anatomically plausible inputs do not necessarily provide incremental predictive value. Task-specific input auditing can improve joint model performance, reduce input redundancy and preserve anatomically traceable predictions. More broadly, this framework provides a testable approach for selecting correlated inputs in medical-imaging systems in which regional views re-present anatomy already contained in the whole image. Keywords: Forensic odontology; Dental age estimation; Sex classification; Panoramic radiography; Multi-input learning; Anatomical input auditing

The automatic detection of thoracic pathologies remains a challenge when the visual evidence present in the chest X-ray (CXR) is subtle, ambiguous, or practically imperceptible, and the available textual information is limited, especially in the early phases of patient care. To address this limitation, in this work a clinically motivated subset of MIMIC-CXR is taken as the study set, composed of four diseases associated with dyspnea presentations in the emergency department (heart failure, pulmonary infection, COPD/asthma, and pulmonary embolism), along with a control class. This selection allows evaluating the multimodal fusion of visual and textual information in a clinically diverse set where the contribution of each modality can vary depending on the considered pathology and the quality of the available information. In this work, a comparative study of different combinations of visual and textual encoders is presented, with the objective of identifying the most suitable configuration for multimodal fusion. Based on this evaluation, we propose Neural Gated Fusion, a GMU-based architecture that incorporates an adapted Gated Multimodal Unit to regulate the contribution of visual and textual representations, in contrast to strategies based on static feature concatenation or late decision-level fusion. This architecture uses a neural gate to dynamically regulate the contribution of the visual representation of the radiograph and the textual representation of the clinical report before the multi-label classification stage. The experiments conducted on 25,245 multimodal pairs show that the ViT + ClinicalBERT combination with modulated fusion achieves a macro ROC-AUC of 0.8545 and a macro F1-score of 0.6528, outperforming unimodal models, decision-level fusion, and static early fusion. The per-class analysis suggests that the designed fusion strategy especially improves performance in pathologies with limited visual evidence, such as pulmonary embolism, without compromising the detection of normal cases. These results suggest that the dynamic regulation between image and clinical context is an effective strategy for improving the robustness of multimodal thoracic pathology detection systems.

The temporal resolution of medical imaging sequences used to drive patient-specific computational fluid dynamics (CFD) simulations remains limited, typically providing 10-20 frames per cardiac cycle. Therefore, temporal interpolation to reconstruct left atrial (LA) wall motion and boundary conditions is required, but its impact on hemodynamic predictions has not been systematically characterized. To investigate this, we constructed high-temporal-resolution reference wall-motion data using electromechanical (EM) simulations on five patient-specific atrial geometries with a history of atrial fibrillation. We then generated temporally downsampled datasets to emulate clinical frame rates (5, 10, 20, and 40 frames per cycle) and performed CFD simulations to isolate the effects of temporal undersampling on hemodynamic metrics. The focus was placed on kinetic energy, KE, and residence time, TR, particularly in the left atrial appendage (LAA), where thrombosis is most likely to occur. We employed an immersed boundary method to prescribe the wall motion and computed blood TR through a passive scalar transport equation. Results indicate that while global LA hemodynamic indices were marginally affected by the frame rate (errors < 9%), LAA metrics were more sensitive with errors up to 31% compared to reference values. The results based on 20 and 40 frames per cycle yielded favorable agreement with reference results, while 5-and 10-frame reconstructions showed larger, though not systematically biased, deviations from the reference. Importantly, patient ranking by blood-stasis indices was largely preserved. The analysis suggests that patientspecific LA reconstructions derived from dynamic CT imaging provide a reliable basis for estimating LAA blood-stasis indices. Higher frame rates ([&ge;] 20 per cycle) offer improved quantitative accuracy, while lower temporal resolutions may remain informative for patient stratification purposes, where relative ranking is more relevant than absolute accuracy.

Postmortem computed tomography (PMCT) is increasingly used in forensic medicine, yet image interpretation remains largely manual and time-consuming, and supervised deep learning is constrained by the scarcity of annotated forensic datasets. This study investigated whether diffusion-based generative models can detect fatal cerebral haemorrhage on PMCT through reconstruction-based anomaly detection, using solely image-level labels derived from the documented cause of death. Postmortem brain CT examinations acquired at a single forensic institute between 2011 and 2025 were retrospectively analyzed, comprising 265 healthy subjects and 33 cases of fatal intracranial haemorrhage. After automated brain extraction, slice selection and intensity normalization, two diffusion frameworks were trained: an unsupervised denoising diffusion probabilistic model (AnoDDPM) trained on healthy anatomy alone, and a weakly supervised classifier-guided diffusion model exploiting image-level labels. Anomaly scores were derived from the reconstruction error between each slice and its pseudo-healthy reconstruction, and detection was assessed at the image level using 95% bootstrap confidence intervals. AnoDDPM achieved an AUROC of 0.928 (95% CI 0.812-0.989) and an AUPRC of 0.794 (95% CI 0.429-0.946), whereas the classifier-guided model reached an AUROC of 0.991 (95% CI 0.984-0.997) with an AUPRC of 0.960 (95% CI 0.855-0.994). These results demonstrate excellent feasibility of diffusion-based anomaly detection for forensic PMCT and indicate that weak, image-level supervision substantially improves detection, thereby supporting the future development of automated triage tools for postmortem brain imaging.

This study emphasizes the potential of computational techniques in cancer risk assessment, lighting opportunities for specific and data-driven healthcare solutions. This study examines the use of artificial intelligence (AI), machine learning (ML), and deep learning (DL) approaches to improve cancer risk assessment using a Kaggle dataset. The study uses Java-based ML software to create and evaluate multiple predictive models, taking advantage of its powerful libraries and frameworks for processing and analyzing cancer risk indicators. This work analyzes model performance using 10-fold cross-validation, resulting in reliable generalization and accuracy estimates. Several classification techniques, such as Random Forest (RF) logistic regression (LR), decision trees (DT), Naive Bayes (NB), and Multi-layer perceptron (MLP), are used to assess their efficacy in predicting risk levels for various cancer types. To measure classification effectiveness, key performance metrics such as accuracy, precision, recall, and F1 score are produced, in addition to multi-class confusion matrices. The results show that the RF model is the best classifier for classification, with accuracy of 99.85%, F-measure of 99.80%, precision of 99.80%, and sensitivity of 99.90%. These findings demonstrate the model's ability to effectively estimate cancer risk levels among individuals. of cancer risk estimations, allowing for earlier discovery and more effective medical care.

Background: Large language models (LLMs) have demonstrated strong performance on standardized medical examinations, with recent studies reporting performance approaching or exceeding that of senior medical residents. However, examination accuracy alone does not establish how models arrive at their answers or the relative contributions of the clinical vignette, imaging, and answer options to model performance. We evaluated the contribution of these question components to LLM performance on the Orthopaedic In-Training Examination (OITE). Methods: We evaluated three open-source Ministral-3 models (3B, 8B, and 14B) and five proprietary models (Claude Haiku-4.5, Sonnet-4.6, Opus-4.8, GPT-5.6 Luna, and GPT-5.6 Terra) using 792 OITE questions from 2020 through 2024, comprising 434 questions containing clinical images and 358 without images. Question components were sequentially removed across five image-containing and three non-image-containing conditions. Accuracy was summarized using the median and IQR and compared using paired analyses with Holm adjustment. Semantic similarity between model-generated explanations and reference discussions was assessed using BioMedBERT. Results: For image-containing questions, pooled accuracy was 59.13% (58.21-60.08) with complete information and 59.13% (58.18-60.11) without images. Accuracy decreased to 49.05% (48.07-50.00) without the clinical vignette, 45.68% (44.70-46.69) without the vignette and images, and 36.49% (35.51-37.44) when the vignette, images, and question were removed. For non-image-containing questions, accuracy decreased from 72.94% (72.10-73.85) with complete clinical context to 53.53% (52.41-54.68) without the vignette and 37.36% (36.28-38.48) with answer options alone. OpenAI GPT-5.6 Terra achieved the highest full-context accuracy in both image-containing 81.80% (80.41-82.95) and non-image-containing 94.13% (93.30-94.97) questions. Removal of images alone did not significantly affect accuracy for any model, whereas removal of the clinical vignette significantly reduced accuracy for most models. When provided with only the answer options, without the question stem, clinical context, or images, the observed accuracy for every model exceeded the 25% expected from random selection. Specifically, the highest accuracy was 45.62% (44.24-47.24) for Claude Opus 4.8 on questions originally containing images and 46.09% (44.41-47.77) for GPT-5.6 Terra on questions without images. Although closed-source models achieved substantially higher accuracy than open-source models, all models performed above chance when provided answer options alone. BioMedBERT similarity remained high despite substantial differences in accuracy and was not significantly associated with accuracy by Spearman correlation ({rho}=-0.17; P=0.29). Conclusion: Clinical vignettes were the primary contributor to improving LLM performance on OITE whereas images had minimal effect. The high accuracy when only answer options were given suggests the models may be learning spurious correlations. Furthermore, BioMedBERT semantic similarity remained high across models despite wide differences in accuracy and did not consistently distinguish correct from incorrect responses. Thus, high examination performance may not directly reflect clinical reasoning, highlighting the importance of assessing how LLMs use the information provided when interpreting benchmark performance.

Background and Objective: Heart rate variability (HRV) analysis provides a non-invasive method for quantifying autonomic modulation from electrocardiographic recordings. However, practical HRV analysis often depends on fragmented workflows, limited signal-quality review, and software tools optimized for either human or preclinical recordings, but not both. This study developed and evaluated HRV-GUI, a MATLAB-based graphical interface for electrocardiogram (ECG)-derived HRV analysis in translational biomedical research. Methods: The HRV-GUI integrates electrocardiographic and RR interval loading, human and rat analysis modes, preprocessing, segment selection, automated R-peak detection, manual peak correction, RR interval generation, multi-domain HRV computation, diagnostic visualization, result export, and session saving/loading. The software was evaluated using deterministic synthetic RR interval datasets, baseline recordings from healthy human controls and healthy rats, and a clinical use-case comparison between healthy controls and patients with diabetic gastroparesis. Results: The HRV-GUI produced expected outputs in synthetic RR validation tests, including constant RR sequences, alternating RR sequences, outlier-containing RR sequences, and low-frequency- or high-frequency-dominant sinusoidal RR modulation. The software generated physiologically plausible HRV profiles in both human and rat recordings. In the clinical use-case analysis, patients with diabetic gastroparesis showed higher heart rate and sympathetic index, together with lower respiratory sinus arrhythmia, absolute low- and high-frequency spectral power, standard deviation of normal-to-normal intervals (SDNN), root mean square of successive differences (RMSSD), percentage of successive RR intervals differing by more than 50 ms (pNN50), Poincare short-term variability (SD1), and Poincare long-term variability (SD2) compared with healthy controls. Conclusions: HRV-GUI provides an integrated biomedical software workflow for ECG-derived HRV analysis. The validation results support its use for controlled RR testing, human and rodent ECG recordings, and clinical autonomic assessment in diabetic gastroparesis.

Purpose: Atrial fibrosis is associated with thromboembolic risk in atrial fibrillation, but how fibrosis-related contractile impairment alters blood transport in the left atrium (LA) and left atrial appendage (LAA) remains unclear. Prior multi-physics simulations coupled electromechanical (EM) and computational fluid dynamics (CFD) models to show that fibrosis impairs LA function and flow, but were limited to booster-only contraction and could not separate the effects of fibrosis burden from its spatial distribution. Here we examined how fibrosis burden and spatial pattern independently relate to long-term LA and LAA hemodynamics in full-cycle, multi-beat simulations on a common LA anatomy. Methods: Three fibrosis patterns derived from late gadolinium enhancement MRI were considered: one from the patient providing the common anatomy and two from additional patients, with the latter mapped onto the common anatomy via universal atrial coordinates. All three patterns were applied globally at matched burdens of 15.6% and 31.2%, yielding six global-fibrosis cases and a non-fibrotic reference. Two patterns were also confined to the LAA at three local burdens (10%, 25%, and 50%), yielding six LAA-fibrosis cases. A 3D EM model for the LA anatomy was coupled to a 0D circulatory model, so that the simulations resolved the reservoir, conduit, and booster phases. Full-cycle EM-derived wall motion drove subsequent multi-beat CFD simulations. Region-specific emptying fraction (EF), non-dimensional kinetic energy (KE), and the 90th percentile of blood residence time (TR) were evaluated for the LA and LAA. Results: Left ventricular stroke volume remained nearly constant across cases, while increasing global fibrosis burden progressively reduced LA and LAA EF (LA: 36% to 30-31%; LAA: 44% to 28-38%) and monotonically increased LA TR across all patterns. Non-dimensional LA KE showed an approximately quadratic association with LA EF across patterns, whereas LAA KE was more scattered but still increased with LAA EF. In contrast, the relationship between the 90th percentile of LAA TR and LAA EF showed pattern-specific, non-monotonic trends. Conclusions: Fibrosis burden and spatial distribution jointly shape LA and LAA hemodynamics, but through different mechanisms. LAA EF is a robust functional descriptor of chamber-averaged kinetic energy regardless of fibrosis pattern, whereas LAA blood stasis depends additionally on the spatial distribution of fibrosis beyond what bulk emptying fraction predicts. These findings support incorporating imaging-derived fibrosis pattern, and not only burden, alongside functional metrics when assessing thrombogenic risk in the LAA.

Purpose: Quantifying aortic morphology is central to surgical planning for thoracic endovascular aortic repair (TEVAR), yet no consensus exists on how best to represent three-dimensional aortic shape for outcome prediction. Two broad strategies have emerged: statistical shape analysis (SSA), which relies on statistical methods and dimensionality reduction to capture the most significant shape modes, and geometrically-informed approaches that extract descriptors grounded in differential geometry. Here, we directly compare these paradigms on a cohort of 290 CTA scans classified by surgical outcome (non-pathological, successful TEVAR, failed TEVAR). Methods: For the geometrically-informed approach, we use a two-dimensional feature space using normalized fluctuation in integrated Gaussian curvature $\widetilde{\delta K}$ and mean aortic radius $R$. For SSA, we construct a point-cloud shape model with dimensionality reduction using Principal Component Analysis (PCA) and evaluate classification performance as a function of the number of retained principal components. Results: SSA's leading principal components encode variations in global aortic size and are statistically redundant with ($R$, $\widetilde{\delta K}$), yet they lack a one-to-one correspondence with interpretable anatomical quantities. Testing on an unseen, independent dataset reveals that the geometrically-informed approach provided better generalizability than SSA. Using Gaussian process classification with 10-fold cross-validation, we find that the geometrically-informed approach achieves a higher weighted $F_1$ score than SSA achieves with up to 20 principal components. While SSA's full-dataset accuracy rises above 90\% with increasing dimensionality, this gain is driven by overfitting rather than genuine discriminative power. Conclusion: These results demonstrate that geometrically-informed descriptors offer a more interpretable, robust, and clinically translatable framework for aortic disease classification than data-driven statistical shape representations.

Purpose: To quantify, in a computational fluid dynamics (CFD) benchmark, how spatial degradation, temporal acquisition and noise change the error of plane-based pulmonary-artery flow endpoints. Methods: Twenty-two historical, non-converged CFD simulations (11 geometry families) were voxelised at 2 mm with 80 frames per cycle. Simulations with numerical divergence were excluded, leaving 13 (7 families). Fields were degraded to 4 and 8 mm and with noise, plane waveforms were sampled at 20 phases with one-interval averaging, and errors were computed against the noise-free reference waveform on 66 geometric planes. Family-mean paired differences are reported with t-based 95% confidence intervals under an analysis plan fixed in advance. Results: Spatial degradation from 2 to 4 mm increased full-waveform nRMSE by 0.0032 [0.0021, 0.0043], in 7 of 7 families; cycle-mean flow bias accounted for 23% of the squared error, against 60% at 8 mm. With SNR scaled to voxel volume, the increase was not resolved (0.0002 [-0.0012, 0.0015]). Twenty-phase acquisition increased PI error (0.0223 [0.0214, 0.0233]) and diastolic-trough timing error (0.63 [0.52, 0.74] frames) more than spatial degradation, and its point estimates for these endpoints were larger with every recovery method; the nRMSE ordering and peak timing depended on the recovery method. The implementation of the downsampling operator mattered: an alternative, corner-aligned implementation of the same 4 mm degradation gave a larger spatial effect (nRMSE difference 0.0010 [0.0002, 0.0017]). Conclusion: In this benchmark, a 2 to 4 mm degradation produced a small, consistent, mostly shape-related nRMSE increase that depended on the noise model, whereas 20-phase acquisition changed PI error and trough timing more. The results describe benchmark behaviour, not clinical acquisition requirements.

Prostate cancer, a leading cause of cancer-related deaths among men globally, necessitates the development of precise diagnostic and treatment strategies. Accurate segmentation of prostate cancer in medical imaging, particularly in MRI scans, is crucial for early diagnosis and clinical decisions. Traditional manual segmentation techniques, while efficient, are labor-intensive and necessitate significant expertise, resulting in an increasing demand for automated alternatives. The RSAUNet is a novel deep learning architecture designed to improve prostate cancer segmentation. It incorporates essential components, including Residual Blocks, Swin Transformer Blocks, and Attention Mechanisms within a U-Net architecture. This markedly enhances the model's capacity to discern complex anatomical features and accurately segment malignant tissues. Using sophisticated deep Learning methods, RSAUNet addresses the complexities of prostate imaging, delivering reliable, consistent segmentation results. The model was evaluated against various cutting-edge techniques on extensive multi-parametric MRI datasets, attaining an impressive Dice coefficient (DC) of 0.998 and a Jaccard Index (IoU) of 0.965. These findings highlight the innovative characteristics of RSAUNet and its capacity to transform prostate cancer diagnosis and treatment strategies. The proposed model surpasses current methods and shows potential for practical clinical applications, providing an efficient, precise, and scalable solution for automated prostate cancer segmentation and fostering optimism for the future of medical imaging and diagnosis.

AO_SCPLOWBSTRACTC_SCPLOWIn this work, we publish the 3D unsteady Navier-Stokes numerical simulations and meshes of coronary arteries reconstructed from invasive X-ray coronary angiograms acquired during in the Fractional Flow Reserve versus Angiography for Multivessel Evaluation 2 (FAME 2) trial. Out of the 914 clinical images, 779 vessels are successfully reconstructed and meshed. The remaining 135 vessels have been discarded since they exhibited self-intersecting geometry during the reconstruction process. The meshes are hexahedral and all of them have the same number of vertices and identical connectivity; their high quality is demonstrated using standard mesh quality indices. The simulations are performed using the Finite Element Method (FEM) with state-of-the-art coronary boundary conditions applied at the outlet. The motivation behind this effort relies from the scarcity of publicly available numerical haemodynamics data, despite the growing interest in data-driven modeling and machine learning techniques. The database is available at the link: https://doi.org/10.7910/DVN/GPCUNS

Despite the use of technological countermeasures, studies show that the sensory conflict, and thus symptoms of motion sickness such as nausea, cannot be completely prevented while traveling as a car passenger. Accordingly, reliable prediction of motion sickness will represent a possible solution for intelligent vehicles to avoid the abortion of the drive, e.g., through adapted driving strategies or route planning. In previous works, the authors were able to demonstrate an application-oriented parameter set for the popular 6DOF-SVC model to estimate motion sickness incidence (MSI) in passengers with a lowered gaze on a non-driving-related task. In the current study, these parameters were validated and further optimized using five previous conducted motion sickness driving experiments. For the first time, it is also shown that the expected natural MSI fluctuations based on the driving dynamic profile, due to route characteristics or driving style, as well as the variation in experimental determined MSI is approximately 10%. Moreover, findings indicate that the participants appear to have an individual MSI threshold that correlates with their susceptibility to motion sickness, however, this threshold fluctuates too strongly across experiments to be suitable for practical application. As a result, the model can successfully estimate motion sickness risk across the test tracks, however, inaccuracies in the predicted MSI trajectories remain.

Esophageal cancer (EC) is characterized by complex transcriptional alterations and therapeutic resistance, posing challenges for traditional computational methods. In this study, we propose a deep learning (DL)-based computational framework to identify important genes and biologically relevant pathways in bulk cell RNA-seq data (GSE234304 and GSE273848), which comprise tumor and non-tumor esophageal tissue samples. A fully connected feedforward neural network was trained for binary classification, and two feature selection strategies were implemented: Neural Network followed by Support Vector Regression (NN+SVR) and Integrated Gradients combined with SVR (IG+SVR). The genes were then ranked according to their weights in SVR deriving the importance scores, and the top 100 genes were subjected to enrichment analysis using Enrichr and Metascape. The proposed DL-based approaches identified a greater number of expressed genes across established esophageal cancer cell lines than LIMMA, SAM, and the t-test did. Specifically, in the GSE234304 dataset, IG + SVR, our best method, identified a total of 9 expressed genes while the best baseline method, LIMMA, identified a total of 3 expressed genes. In terms of GSE273848 dataset, IG + SVR was also the best identifying a total of 11 expressed genes while the best baseline method, t-test, had a total of 7 expressed genes. The key genes identified included CEBPB, SUMO1, RORA, STAT1, GATA, OCT1, RUNX1, and NR3C1, as well as pathways related to nucleoprotein maturation, collagen fibril organization, the immunoglobulin-mediated immune response, immune regulation, and insulin signaling. These results show that combining neural networks and attribution-based regression creates an effective and interpretable framework for selecting genes in esophageal cancer research.

Background: Regurgitant valvular heart disease (rVHD) is a major cause of cardiovascular morbidity. Echocardiography is the diagnostic standard but is resource-intensive for large-scale screening. Electrocardiography (ECG) has shown promise for predicting incident rVHD, yet performance varies across phenotypes, particularly for aortic regurgitation (AR). Chest radiography (CXR) provides complementary structural and hemodynamic information. We hypothesized that a multimodal model integrating ECG and CXR would improve prediction of incident moderate-to-severe rVHD. Methods: In this retrospective multicenter study, we identified 212,888 paired ECG-CXR examinations from 116,380 patients across two Chinese centers. Baseline ECG and CXR were obtained within 60 days of echocardiography. Outcome was progression to moderate-to-severe AR, mitral regurgitation (MR), or tricuspid regurgitation (TR). We developed a multimodal neural network with pretrained unimodal encoders, token-level cross-modal fusion, and a class-specific gating mechanism that adaptively weighted ECG-only, CXR-only, and fused predictions. Performance was assessed using C-index, AUROC, AUPRC, decision curve analysis, net reclassification improvement (NRI), and Kaplan-Meier stratification. Results: Multimodal fusion consistently outperformed unimodal models across all phenotypes. For AR, C-index improved from 0.616 (ECG-only) to 0.713 (multimodal; AUROC 0.729, AUPRC 0.972). For MR, multimodal C-index was 0.801 (AUROC 0.814, AUPRC 0.972), versus 0.782 for ECG and 0.775 for CXR alone. For TR, multimodal and CXR-only models showed similar discrimination (C-index 0.802), but multimodal fusion yielded greater net benefit on decision curve analysis. NRI was positive across all time horizons (1-5 years) for all valve types. Grad-CAM interpretability analyses revealed that ECG attention localized to leads II, V-V (AR), leads I, II, aVF, V-V (MR), and inferior/right precordial leads (TR); CXR attention highlighted chamber-specific enlargement and pulmonary congestion patterns consistent with pathophysiology. Conclusion: A multimodal deep learning model integrating ECG and CXR significantly improved prediction of incident rVHD compared with ECG alone, with the greatest benefit observed for AR. The model leveraged complementary electrical and structural information, demonstrated biological plausibility through interpretability analyses, and provided consistent clinical utility. Given the widespread availability and low cost of both modalities, this approach offers a scalable tool for risk stratification in routine care. Prospective studies are warranted to validate clinical implementation.

Background: Spinal cord injury (SCI) is frequently associated with orthostatic hypotension, defined by a sustained decrease in blood pressure upon assuming an upright posture due to impaired autonomic regulation. Cardiovascular spinal cord epidural stimulation (CV-scES) can regulate systolic blood pressure (SBP) in people with SCI, but stimulation paradigms are highly individualized. To make this treatment available to more patients, we developed an algorithm to tailor individualized CV-scES paradigms that closely mimic researcher-developed paradigms. Methods: We performed an offline analysis using datasets collected from eight individuals with SCI with epidural stimulators implanted over the lumbosacral spinal segments. During data collection, researchers modulated stimulation parameters with the goal of maintaining SBP between 110-120 mmHg. Each two-hour dataset included synchronized SBP and stimulation recordings. We ran optimization analyses offline to determine temporal requirements before modifying stimulation amplitude to mitigate out-of-range SBP. Results: The algorithm parameters that best matched researcher-selected stimulation changed relatively quickly during the first 12 min (one every ~40 sec), and more slowly thereafter (one every ~79 sec). Overall, algorithmic stimulation closely tracked researcher-controlled stimulation, with a mean correlation coefficient of 0.94. To evaluate online performance, we tested the algorithm in real time with a single participant. We found that a faster approach was needed to respond to changes in SBP caused by rapid, unpredictable events, such as postural changes. We implemented a sigmoid-based paradigm that determined the time to wait before changing stimulation as a function of the current SBP, with worse SBP values requiring faster responses. The new paradigm outperformed the original algorithm and researcher-controlled stimulation across measures of SBP stability, though recovery from a postural tilt maneuver remained slower than with researcher control. Conclusions: Our results indicate that algorithmic stimulation may minimize assistance required from researchers and participants, making CV-scES more feasible for clinical translation.

Acute myocardial infarction (AMI) is one of the leading cardiovascular diseases worldwide and remains a major cause of mortality. Early risk prediction can help clinicians identify high risk patients shortly after admission and support timely monitoring and individualized treatment. Previous AMI risk assessment approaches predominantly rely on a single model structure or fixed feature representation, which may limit their ability to capture diverse risk related patterns and reduce predictive performance. To address these challenges, we propose MetaAMI, a random projection based meta-learning framework for AMI outcome prediction. Specifically, patients were selected based on ICD-9 and ICD-10 diagnostic codes for AMI from the Medical Information Mart for Intensive Care IV (MIMIC-IV) v3.1 database. Features were transformed through multiple random projections, with each random projection generating a distinct lower dimensional feature representation. Subsequently, baseline classifiers were trained on each lower dimension representation to predict in-hospital mortality among AMI patients. The predictions were aggregated to construct an integrated feature representation, which was used as input to a meta-learner architecture. By effectively integrating complementary information from diverse baseline models, the meta-learner refined the decision boundary and enhanced overall predictive performance. Survival analysis and SHapley Additive exPlanations (SHAP) analysis were further performed to evaluate clinical utility and interpret the model predictions. Benchmarking results based on MIMIC-IV dataset suggested that our MetaAMI consistently outperformed all the baseline classifiers across seven evaluation metrics including Accuracy, Area Under the Curve (AUC), F1 Score, G-Measure, Jaccard Index, Youden J, and Matthews Correlation Coefficient (MCC). In addition, feature importance analysis showed clinically relevant predictors of in-hospital mortality. In summary, MetaAMI provides an effective and robust solution for machine learning based AMI risk prediction. We anticipate that the application of MetaAMI will have a positive impact on clinical risk stratification and personalized treatment strategies for AMI.

Depression in Alzheimer's disease (AD) is clinically challenging to characterise because affective and cognitive manifestations frequently overlap with the underlying neurodegenerative condition. This study investigates depression-related speech patterns in AD using an explainable multimodal framework that integrates linguistic and acoustic features from elicited picture-description speech. To address the absence of directly observed depression labels in the target AD cohort, we employ a source-informed cross-dataset transfer strategy. Unsupervised clustering identified two distinct speech sub-phenotypes within the AD cohort. By comparing these unsupervised representations to a depression-labelled source corpus, one cluster demonstrated mathematical proximity (measured via cosine similarity in a shared latent space) to the source depression centroid. We explicitly acknowledge that this proximity reflects geometric similarity in a shared latent space, which may encapsulate overlapping domain characteristics alongside affective variance. To validate the structural integrity and separability of these discovered clusters, a proxy supervised classification task was employed. The retained multimodal artificial neural network (ANN) integrated information from both modalities and achieved 92.73\% cluster separability accuracy. This demonstrates that the representation-derived two-cluster partition was highly learnable on the independent internal hold-out test set; it does not establish clinical validity, longitudinal stability, or diagnostic validity of the identified sub-phenotypes. For interpretability, Shapley Additive exPlanations (SHAP) analysis was utilised to extract feature attributions and organise influential acoustic and linguistic features into a multimodal taxonomy. Furthermore, a Multimodal SHAP Attribution Magnitude (MSAM) score was derived as an exploratory model-based attribution score rather than a clinically validated severity measure. Ultimately, this study provides an interpretable computational baseline for characterising depression-related speech variance in AD, establishing a structural foundation for future prospective validation using independently labelled clinical cohorts.