MOTIVATION:Network-based drug repurposing leverages the principle that drug effects propagate through protein-protein interaction (PPI) networks similarly to disease perturbations. Graph-based embedding methods exploit this principle by learning low-dimensional representations that capture the network neighborhood of disease and drug target proteins, enabling similarity-based prioritization of therapeutic candidates. However, existing approaches that learn fixed embeddings from a static graph cannot accommodate perturbations to the graph structure without costly recomputation or retraining, making them impractical for large-scale screening where each drug generates a distinct transcriptional perturbation. OBJECTIVE:We present GraphPert, a Graph Autoencoder (GAE) framework that propagates transcriptional signatures through the human PPI to enable efficient network-based virtual screening. METHODS:Transcriptional signatures from the Connectivity Map L1000 platform are encoded as symmetric edge-weight modifications in the PPI, scaling each interaction by the expression changes of both participating proteins. A GAE pre-trained on the human PPI (N=12,458 proteins) produces latent displacements for each perturbation, and compounds are ranked by cosine similarity to a therapeutically relevant reference-the BRAF V600E knockout (KO) in A375 melanoma cells. RESULTS:In retrospective screening against a decoy library of known MAPK cascade inhibitors seeded among compounds without known MAPK pathway activity (1:9 ratio), GraphPert achieves AUC >0.80. Notably, while raw signature-based screening recovers predominantly compounds with targets proximal to the MAPK pathway, GraphPert additionally identifies compounds with more distal targets-including HDAC, proteasome, and PARP inhibitors-that converge on the same downstream functional modules suppressed by BRAF KO. CONCLUSION:GraphPert demonstrates that propagating transcriptional signatures through a pre-trained GAE extends the reach of virtual screening beyond direct transcriptional similarity, enabling the discovery of mechanistically diverse compounds that recapitulate the network-level effects of a therapeutic reference perturbation.
Biomarker discovery, essential in neuroscience research, has traditionally been based on static molecular omics, which frequently reveal established pathogenesis at the network level when irreversible structural damage occurs. In this review, the paradigm shift of functional "digital phenotyping" is discussed, along with its implementation by combining human induced pluripotent stem cell (hiPSC)-derived brain organoids with high-density multielectrode arrays (HD-MEAs). Researchers can use advanced architecture deep learning (DL) algorithms, such as Convolutional and Graph Neural Networks, to identify predictive signatures of diseases at the network level by separating complex spatiotemporal dynamics in firing patterns. These digital functional biomarkers can their ability to detect abnormal burst kinetics, abnormal oscillatory coupling, and abnormal functional connectomes before the onset of structural cell loss. Moreover, we discuss the underlying methods of artificial intelligence (AI) analysis and advances in phenotypic drug discovery using organoid electrophysiology, changing the paradigm of therapeutic targeting from the clearance of molecular aggregates to the functional recuperation of neural circuit health. In conclusion, a combination of powerful 3D culture technology, state-of-the-art microelectronics, and machine learning offers an extensive novel translational platform for dissecting early neurodevelopmental and neurodegenerative diseases, ultimately driving precision therapeutic approaches for the brain and other organs.
Parkinson's disease (PD) is a progressive neurodegenerative disorder with a complex etiology, in which genetic variability in non-coding regions and epigenetic mechanisms play a central role. This study investigated allele-specific expression (ASE) patterns in peripheral blood RNA-Seq data from 487 individuals with idiopathic PD (iPD) and healthy controls from the PPMI initiative, aiming to characterize disease-associated transcriptional imbalances and their relationship with CpG islands. We identified 31,219 significant variants at heterozygous sites, with 4537 variants unique to the iPD group (14.53% of the total) mapping to 921 genes, a number significantly higher than that observed in controls (73 genes). Functional analysis revealed consistent enrichment in immune pathways (HLA-B antigen presentation, FcγR signaling, Th17 differentiation) and processes involving ubiquitination and protein degradation, suggesting an integrated molecular signature of immune dysregulation and proteostatic stress. Chromosomal distribution showed hotspots on chromosomes 1, 2, 6, 12, and 17, with a predominance of transition variants in CpG-rich regions. However, no direct linear correlation was observed between variant characteristics and CpG island metrics (density, GC content), indicating that regulatory effects may operate independently of local methylation. Protein interaction network analysis highlighted sub-networks featuring hub genes such as PINK1, GBA1, LRRK2, and PRKN, connected to autophagy and immune response pathways. These results support the view that ASE captures a broad regulatory signature in iPD, distinct from the DNA methylation context, and reinforces the role of immunogenetic dysregulation as a primary component of the disease's pathophysiology, offering new targets for mechanistic investigation and peripheral biomarkers.
BACKGROUND:Excessive adiposity drives cardiometabolic morbidity through macroscopic systemic mechanisms rather than isolated metabolic defects. While conventional statistical models effectively estimate marginal biomarker associations, they cannot resolve the directed and cyclic regulatory dependencies underlying this pathogenesis. Network physiology offers a robust mathematical framework to delineate these complex, feedback-driven interactions. METHODS:Applying the idopNetwork framework to cross-sectional data from a primary clinical cohort (N=7,343) and an external validation cohort (N=6,881), we reconstructed directed, weighted hematometabolic networks across four body mass index (BMI) strata. This analytical pipeline mapped static clinical snapshots onto a continuous coordinate space to extract quasi-dynamic autoregressive trajectories. GLMY path homology was subsequently integrated to quantify macroscopic network rigidity via higher-order topological features. RESULTS:Progressive adiposity manifested a graded topological reorganization along the allostatic load gradient, characterized by the functional polarity inversion of core physiological hubs. Uric acid initially centralized as a primary conduit of metabolic load, preceding the transition of hemoglobin from a homeostatic anchor into a predictive promoter of systemic inflammation. GLMY path homology defined advanced obesity as a rigid systemic gridlock, quantified by the abnormal accumulation of persistent one-dimensional cyclic loops (β1) and two-dimensional voids (β2). This terminal architecture exhibited profound sexual dimorphism. Male adiposity degraded into hyper-reactive inflammatory β1 cyclic tangling, whereas the female network experienced a near-complete depletion of preexisting β2 voids upon entering advanced obesity, culminating in extensive topological barrenness and the attenuation of compensatory feedback. CONCLUSIONS:The internal physiological ecosystem structurally deteriorates from robust homeostatic buffering into an advanced, sexually dimorphic allostatic deadlock. By decoupling directed regulatory flows from systemic confounding, this macroscopic algebraic approach challenges universal treatment protocols, highlighting the necessity for sex-specific precision interventions in managing obesity-driven cardiometabolic risk.
Cancer progression is driven by a tumor microenvironment (TME) that suppresses anti-tumor immunity through structural, cellular, and metabolic barriers. We developed a computational model to compare the relative impact of four major suppressive mechanisms: extracellular matrix (ECM) remodeling, PD-1-mediated checkpoint inhibition, regulatory T cell (Treg) suppression, and glucose-related metabolic limitation. Model parameters were calibrated using experimental data to ensure biological fidelity and predictive accuracy. Simulations reveal that ECM-mediated immune exclusion is the dominant barrier to anti-tumor activity, followed by PD-1 signaling, Treg suppression, and glucose deprivation. Physical restriction of immune cell access to the tumor represents the primary bottleneck, with other suppressive mechanisms exerting stronger effects only after immune cells have already infiltrated the tumor niche. Interventions targeting ECM normalization produced the greatest reduction in tumor burden and were associated with increased cytotoxic T cells and mature dendritic cells, indicating a shift toward a more immunostimulatory microenvironment. Combination strategies involving ECM targeting and checkpoint blockade further enhanced anti-tumor effects, supporting the idea that relieving stromal exclusion can improve the effectiveness of downstream immunotherapies. Overall, our findings highlight the layered and cooperative nature of TME-mediated immune evasion and suggest that therapies aimed at disrupting ECM-driven immune exclusion may provide the most effective entry point for restoring anti-tumor immunity in solid tumors. These results offer a mechanistic rationale for prioritizing ECM-normalizing strategies in combination with checkpoint inhibition.
BACKGROUND:Poly(methyl methacrylate) (PMMA) is widely used in dental and craniofacial applications; however, its clinical performance is limited by poor surface wettability, moderate mechanical strength, and restricted biological activity. Integrating nanomaterial engineering with computational biology offers an opportunity to better understand biomaterial-cell interactions and support the rational design of functional biomaterials. METHODS:Nickel oxide (NiO) nanoparticles were synthesized via chemical precipitation and incorporated into PMMA to fabricate nanocomposites. Physicochemical characterization included contact angle measurements, Fourier-transform infrared spectroscopy (FTIR), scanning electron microscopy (SEM), energy-dispersive X-ray spectroscopy (EDX), and Vickers hardness testing. Biocompatibility was evaluated using zebrafish embryo developmental assays. To explore biological processes potentially associated with biomaterial-cell interactions, bioinformatics analyses including Gene Ontology (GO), Kyoto Encyclopedia of Genes and Genomes (KEGG), and STRING protein-protein interaction (PPI) network analyses were performed. RESULTS:Incorporation of NiO nanoparticles improved the surface and mechanical properties of PMMA, reducing the contact angle from 105.35° to 90.46° and increasing Vickers hardness compared with unmodified PMMA. Structural and morphological analyses confirmed successful synthesis and homogeneous nanoparticle incorporation. Zebrafish embryo studies demonstrated minimal developmental toxicity, supporting the biocompatibility of the nanocomposite. Bioinformatics analyses identified significant enrichment of pathways related to extracellular matrix organization, cell adhesion, focal adhesion, PI3K-Akt signaling, and oxidative stress regulation. Protein-protein interaction analysis revealed highly interconnected networks associated with ECM-integrin signaling and redox homeostasis, highlighting biological processes potentially associated with biomaterial-cell communication. CONCLUSIONS:PMMA/NiO nanocomposites exhibited improved physicochemical performance and favorable biocompatibility characteristics. The integration of experimental characterization with bioinformatics and network-based analyses provides a systems-level perspective on biomaterial-associated cellular processes and identifies ECM-integrin signaling and oxidative stress-related pathways as candidate biological processes for future experimental validation. These findings support the continued development of PMMA/NiO nanocomposites for oral and craniofacial biomedical applications.
The Transjugular Intrahepatic Portosystemic Shunt (TIPS) is a well-established treatment for complications of portal hypertension in liver cirrhosis, effectively reducing the portal pressure gradient (PPG) and improving transplant-free survival. However, excessive shunting, particularly with larger graft diameters, may increase systemic ammonia levels, predisposing patients to hepatic encephalopathy, and may also precipitate cardiac complications. Computational modelling may therefore eventually serve as a predictive tool for patient selection and pre-procedural planning. In this study, patient-specific TIPS geometries were segmented from CT images and analyzed using computational fluid dynamics with physiologically grounded boundary conditions, before and after TIPS placement. Liver volumetry, hematocrit-dependent blood viscosity, and clinically measured pre-TIPS pressure gradients were incorporated within a coupled 3D-0D multiscale framework. We systematically evaluated the influence of shunt diameter (6-10 mm), puncture location (right, left portal branches and portal bifurcation), angulation, inflow distribution, and outflow partitioning on portal hemodynamics. Results show that the hemodynamic response varies substantially with shunt configuration and patient-specific parameters. Modulating diameter produces a clear trade-off between portal decompression and overshunting. Shunt position, length, and angulation influence PPG only mildly through changes in effective shunt resistance and alter the proportion of TIPS flow originating from the superior mesenteric vein by not more than 10% across configurations. Furthermore, hepatofugal and hepatopetal flow states are mechanistically reproduced, and are governed by the balance between TIPS, intrahepatic portal and sinusoidal-hepatic venous resistances. Overall, this study provides an integrated and physiologically informed framework for understanding and optimizing TIPS configuration in a patient-specific manner.
PURPOSE:There is few research on which genes play an important role in tumors without lymph metastasis. This study aimed to identify candidate molecular alterations preferentially associated with N0-stage LUSC. METHODS:we conducted a comprehensive bioinformatics analysis using publicly available The Cancer Genome Atlas (TCGA) data. Differentially expressed genes (DEGs) were identified separately by comparing N0 tumors and N+ tumors with normal lung tissues. Genes dysregulated in both N0 and N+ tumors were excluded to identify candidate N0-associated genes PPI networks were constructed using STRING and Cytoscape, with module analysis performed via MCODE. Hub genes were identified using multiple Cytohubba algorithms. Functional enrichment analyses were conducted using GO, and KEGG pathways using DAVID. Gene interaction networks were further explored using GeneMANIA. Immune cell infiltration was evaluated with TIMER. Associations with pathological stage and patient survival were assessed using GEPIA and other relevant tools. RESULTS:A total of 1103 candidate N0-associated DEGs were identified, including 748 upregulated and 355 downregulated genes. The PPI network contained five major MCODE clusters. One cluster (MCODE 4) included TTC30A, TTC30B, BBS7, and KIF3B genes implicated in ciliogenesis. TTC30B showed significant differential expression across pathological stages in the overall LUSC cohort. Seven consensus hub genes (ERBB2, CHUK, CASP8, NOTCH1, HNF4A, CREBBP, and IRS1) were identified based on their consistent ranking across multiple CytoHubba algorithms. Upregulated candidate N0-associated genes were primarily enriched in immune-related processes, including B-cell-mediated immunity and humoral responses, whereas downregulated genes were enriched in lysosomal and trans-Golgi network-related pathways. Exploratory immune infiltration analyses identified associations between the four ciliogenesis-related genes and several immune cell populations. CONCLUSIONS:This study identified candidate molecular signatures preferentially associated with N0-stage LUSC, including ciliogenesis-related genes and consensus hub genes. These findings provide hypotheses regarding molecular features of N0-stage LUSC and warrant further validation in independent cohorts and experimental studies.
Accurate visual motion estimation requires the integration of local component-motion signals into a global, pattern-consistent representation. Unikinetic plaids provide a controlled way to probe this process because a drifting component supplies motion energy while a static oriented component constrains the global solution. This study investigated whether short-latency ocular-following responses (OFRs) in children show the temporal sequence predicted by adult and non-human-primate work. Eight typically developing children (mean age 10.5 years; range 5-16 years) were tested using a pediatric-oriented, non-invasive, high-resolution video-oculography system with marker-based head-motion stabilization. Four unikinetic plaid conditions were presented in a brief protocol, and horizontal and vertical eye displacements were estimated from sparsely sampled image pairs during the first 200 ms after motion onset. A component-sensitive horizontal contrast was evaluated over 80-160 ms, and a pattern-sensitive vertical contrast over 100-200 ms. Bootstrap inference was performed within subject. The component-sensitive response was significant in seven of eight children, the pattern-sensitive response in six of eight, and both effects were present in six of eight participants. These findings provide the first pediatric evidence, to our knowledge, that OFRs to unikinetic plaids can reveal an early component-driven response followed by a later pattern-consistent deviation. The results support the feasibility of marker-stabilized video OFR measurements as a low-demand biomedical signal for studying motion-integration dynamics in children and motivate larger normative and clinical validation studies.
Atrial fibrillation (AF) poses significant challenges for effective treatment, particularly in persistent AF cases (psAF). Current pharmacological and ablative therapies remain suboptimal due to the complex nature of arrhythmogenic substrates. In this context, patient-specific simulations have emerged as a promising tool to improve therapy planning and personalization. In this study, we developed atrial digital twins calibrated to reproduce clinical activation patterns and evaluated their ability to reproduce clinical trends in biomarkers derived from intracardiac electrograms (EGMs).We calibrated patient-specific models for 20 patients with psAF using anatomical meshes derived from computed tomography scans (CT), including fiber direction, atlas-derived tissue heterogeneity, and fibrotic regions estimated from clinical bipolar EGM voltage maps. Electrophysiological properties were personalized through a multi-step calibration procedure including global and local diffusion adjustment and ionic channel remodeling optimization.The calibration workflow significantly reduced discrepancies between simulated and clinical local activation time (LAT) maps obtained after the third extrastimulus of the triple short-coupled extrastimuli pacing protocol (3-Extra), decreasing the mean absolute error (MAE) from 36.3 ± 11.1 ms at baseline to 14.3 ± 3.8 ms after complete calibration (p < 0.01). Total depolarization time (TDT) errors were reduced from 41.3 ± 21.3 ms to 4.9 ± 5.8 ms, indicating substantial correction of baseline conduction discrepancies. Biomarkers derived from simulated EGMs reproduced the clinical trends in activation duration, EGM fractionation and LAT variability between clinically annotated healthy and abnormal atrial tissue (p < 0.01). However, the magnitude of these differences was smaller in the simulations and voltage-related behavior was not accurately reproduced.These results demonstrate the feasibility of generating personalized atrial digital twins that reproduce patient-specific activation patterns and partially capture activation-based EGM biomarker trends. The calibration of these models supports their potential use as simulation tools to investigate atrial conduction abnormalities and potentially guide future therapy-planning strategies, while highlighting the need for improved modeling of voltage-related mechanisms before simulated EGMs can fully reproduce clinical signal amplitudes.
Multimodal physiological signal fusion-particularly electroencephalography (EEG) and electrocardiography (ECG)-is widely assumed to improve emotion recognition over unimodal approaches. Yet the conditions under which fusion genuinely helps, and which fusion architecture is most robust, remain poorly characterised. We present a large-scale ablation study comparing five fusion strategies (EEG-only, ECG-only, concatenation, bidirectional cross-attention, and adaptive gated fusion) on the emotion datasets DREAMER and AMIGOS, and validate ECG signal-quality effects on the stress dataset WESAD. Using a lightweight deep learning architecture (PMMR-DL v3, ∼1.3M parameters) and leave-one-subject-out cross-validation (LOSO-CV) repeated over three random seeds, we evaluate macro F1-score and accuracy for binary emotion classification. Our results reveal pronounced dataset-dependent modality dominance: ECG dominates on DREAMER (F1 = 42.1% vs. EEG 28.6%), whereas EEG dominates on AMIGOS (F1 = 63.8% vs. ECG 13.8%). Critically, multimodal fusion does not consistently outperform the best unimodal baseline. On DREAMER, the strongest fusion strategy (adaptive gated, F1 = 37.1%) still underperforms ECG-only. On AMIGOS, all fusion conditions collapse toward the poor ECG performance (F1 17-21%) because the ECG recordings are noise-dominated. The adaptive gated strategy achieves the lowest inter-seed variance (std = 0.5 on DREAMER), making it the most reliable fusion mechanism when signal quality is adequate. These findings caution against blanket assumptions that multimodal fusion is universally beneficial, and provide practical guidance for modality selection in affective computing pipelines.
Traditional systems of ocular disease diagnosis and many deep learning-based systems are limited in their practice to analyzing fundus images from a single eye, often using raw, unprocessed data. This fails to include the critical correlation between bilateral eyes, has several issues such as low image contrast, noise, and class imbalance, and generally concentrates on the detection of no more than a single disease. In order to bridge the gap, Enhanced Diabetic Retinopathy Detection through Multimodal Integration of Fundus Imaging Features and Clinical Demographic Information utilizing Verifiable Convolutional Neural Network (DD-FIF-CDI-VCNN) is proposed. In the initial phase, the fundus images of the eyes along with their corresponding demographic information are sourced from the Ocular Disease Recognition, which acts as the main feed for the proposed framework. During the pre-processing stage, it uses Robust Consensus Tobit Kalman Filtering (RCTKF) to resize the fundus images, normalizes the pixel intensities, and reduces noise in the process of quality improvement for subsequent feature extraction. The pre-processed fundus images are provided to the ResNet-fused External Attention Network (ResfEANet) based feature extraction that extracts rich visual and texture features from the retinal images. In parallel to this, the demographic attributes are fed to a feature extraction module using TabNet which converts the structured affected person facts into a meaningful function vector. This aims at making sure that relevant patient-specific styles are encoded and prepared for integration with visual functions. Then, the visible and texture capabilities extracted from the fundus photos are combined with the demographic attributes the usage of Hierarchical Multi-Scale Feature Fusion (HMSFF). This creates a complete representation that leverages both imaging and structured affected person statistics. This blended feature set is the input to the categorization model for correct diabetic retinopathy detection. Next, this fused characteristic vector, combining the features extracted from the fundus images with their corresponding demographic attributes, is passed to a Verifiable Convolutional Neural Network (VCNN)-based classifier model. This comprehensive multi-modal representation ensues, rendering the VCNN model capable of accurate detection of diabetic retinopathy classifying normal and diabetes, thereby enhancing predictability and robustness compared with the unimodal approach. The VCNN parameters are fine-tuned by the Warthog Optimization Algorithm (WOA) to ensure improved convergence and better detection performance. The proposed DD-FIF-CDI-VCNN method is analyzed under performance metrics: accuracy, precision, recall, F1-Score, Area under Curve (AUC) and Error rate when compared with existing models.
Cardiovascular risk prediction from heterogeneous physiological signals supports early warning in bedside and wearable monitoring, where ECG, PPG, HRV, signal-quality indicators, activity context, and clinical metadata jointly carry evidence of short-term deterioration. However, existing approaches face two limitations: HRV-based models are physiologically interpretable yet cannot represent waveform morphology, whereas conventional multimodal deep models entangle autonomic regulation, cardiovascular morphology, activity-related interference, patient-specific baseline, and acquisition uncertainty in one latent space. We propose AM-DiMNet, an HRV-anchored autonomic-morphological structured factorization framework predicting clinically recorded cardiovascular deterioration within 24 h. In the primary end-to-end cohort, it integrates ECG, PPG, explicit HRV features, movement- and waveform-instability proxies, clinical metadata, modality-availability masks, and signal-quality scores, then factorizes the fused representation into autonomic, morphological, activity, baseline, and noise components via HRV anchoring and quality-aware fusion. Because no public dataset jointly provides all modalities with longitudinal outcomes, PCG is treated as an architecturally compatible optional branch and is assessed only through task-specific component-level morphology experiments, not as an empirically validated contributor to the primary 24-h endpoint. AM-DiMNet attained an AUROC of 0.895, AUPRC of 0.684, Macro-F1 of 0.804, MCC of 0.641, and ECE of 0.036, raising AUROC from 0.786 and AUPRC from 0.503 over HRV-based XGBoost and surpassing attention and modality-dropout fusion in discrimination, calibration, and incomplete-input robustness. The prediction target was a retrospective composite of time-stamped physiological adverse events and care-process-mediated interventions recorded under the clinical practices represented in the source cohort, and should not be interpreted as a policy-invariant estimate of untreated biological deterioration. These results show HRV can act as an interpretable autonomic anchor for calibrated, physiologically structured risk modeling; the latent-factor analyses offer internal, mechanism-consistent evidence rather than causal or clinician-validated explanations.
Background Vasodilatory endothelial dysfunction (VED) is identified as a precursor to cardiovascular diseases (CVDs) and atherosclerosis. Therefore, early detection of VED is important to prevent the development of severe CVDs and prescribe lifestyle changes or make clinical interventions. Though Acetylcholine-induced coronary vasodilation test is the invasive gold standard to detect VED, specialized devices have provided alternative means of detecting VED non-invasively. With these, they have already detected presence of VED in patients who are already diagnosed with CVD or Diabetes Mellitus (DM). This study focused on early detection of VED in preclinical stages: individuals with pre-diabetes (PDM) and asymptomatic individuals with cardiovascular risk factors (ARF) along with individuals with DM, individuals with CVD with/without DM and healthy (control). Non-invasive physiological parameters were employed to assess VED in these five subject categories. Methods We developed a custom-made device to record non-invasive physiological parameters of digital body temperature, peripheral arterial tone, photo plethysmography and peripheral bio-impedance from the distal end of upper limbs. Multiple indices which represent unique features of these parameters were calculated considering both long-term and short-term variations. Using these, the performance of this device was validated against the flow-mediated dilation (FMD) procedure in a preliminary study with 10 volunteers (age range: 35 - 60 years): 5 healthy subjects (control) and 5 CVD patients with/without DM. Then the clinical study was conducted with 69 participants (age range: 35-60 years) with all the subject categories: healthy (n = 9), ARF (n = 15), PDM (n = 6), type-2 DM (n = 15), 24 CVD with/without DM (n = 24). CVD risk factors identified for the ARF category were physical inactivity, unhealthy diet, smoking, alcohol consumption, body mass index and family history of CVD. Participants with type-1 DM, liver cirrhosis, renal failure, thyroid disease, spinal cord injuries and finger deformities were excluded from the study. Results In the preliminary study, an unpaired t-test and Bland Altman analysis were performed between FMD readings and indices derived from parameters measured from the custom-made device, to validate the device's performance. In the clinical study, the ANOVA analysis showed promising results (p < 0.05) in distinguishing long-term trend variations and short-term variations of VED-associated physiological parameters in test subjects diagnosed with CVD or DM along with preclinical stages (ARF and PDM) compared to the healthy (control). Conclusion We have demonstrated that monitoring VED-associated physiological parameters could potentially enable the early detection of cardiovascular disease (CVD) during its preclinical stages. Thereby, preventive measures or treatments can be initiated at early stages while preventing the severe development of CVDs.
Diabetic Retinopathy (DR) affects millions of people worldwide, but screening at population-scale is still limited by the lack of specialists and the clinical inefficiency of existing deep learning systems, which either perform binary referral or severity grading but not both within a single unified routed architecture. We present a shared EfficientNet-B4 backbone with two independently activatable heads: a binary screener (Head 1) and a joint five-class DR severity and three-class Diabetic Macular Edema (DME) risk grader (Head 2), trained by way of a five-stage progressive curriculum which prevents the gradients of one pathway from interfering with the other. Clinical routing only invokes the grader for referred cases, reducing computational demand as a function of DR prevalence. The screener achieves AUC =0.9789 on APTOS and AUC =0.8685 on zero-shot external validation (Messidor-2); a small calibration sample of 50 Messidor-2 images recovers ≥90% sensitivity, demonstrating that the AUC gap reflects threshold shift rather than a model failure. The grader achieves five-class DR severity QWK =0.7731 with simultaneous three-class DME risk output; an ordinal loss and MixUp remediation stage (Stage 5b) rebalances per-class performance, substantially improving No DR and Severe NPDR F1 at the cost of a partial reduction in PDR F1 and a decrease in Grade 1 F1 from 0.1818 to 0.1429. The principal limitation is that the grader was trained on only 413 IDRiD images, and Grade 1 F1 remains low across all variants, reflecting the scarcity of annotated grading data.
Epilepsy affects roughly 50 million people worldwide and is diagnosed primarily through electroencephalography (EEG), yet the manual review on which this diagnosis depends remains slow, subjective, and inconsistent across readers. A rapidly expanding literature reframes EEG signals as two-dimensional images, allowing ImageNet-pretrained vision backbones to be repurposed for seizure recognition via transfer learning, but no prior work has disentangled the competing tabular-to-image encodings from the choice of architectural family, nor characterized their joint behavior within a single leakage-controlled protocol. The study closes this gap by factorially screening seven encodings (DeepInsight, IGTD, REFINED, Gramian Angular Field (GAF), Markov Transition Field, Tab2Img, and an HSV channel scheme) against ten backbones spanning the convolutional, pure-transformer, and hybrid families across binary, three-class, and five-class tasks from the Bonn EEG corpus. Evaluation employed GroupKFold partitioning, dual seeds, Optuna hyperparameter optimization, paired non-parametric inference with Holm-Bonferroni correction, Cohen's d effect sizes, and Grad-CAM/Attention-Rollout explanations. Across more than 600 final runs, DeiT-Small with IGTD attained a macro-F1 of 0.9868 ± 0.0070 on the saturated binary task, whereas the hybrid CoAtNet-0 with GAF dominated the harder three- and five-class settings (0.9538 and 0.7799). Two reciprocal within-superset confusions jointly accounted for 67.5% of the residual five-class error, while hybrids achieved Cohen's d advantages of +1.2 to +2.0 and occupied the accuracy-efficiency Pareto frontier. Architecture and encoding interact non-trivially with task granularity, and the persistent five-class gap is best closed by enriching the input representation rather than enlarging the backbone.
Pulmonary biomechanical behavior is a critical determinant of the structure-function relationship in the lung under normal and pathological conditions. Pulmonary surfactant plays a key structural role in lung function and significantly contributes to the mechanical response of the lungs during respiration. Additionally, the viscoelastic behavior of the lung parenchyma contributes to the lung's mechanical function and may be altered in disease. However, the contributions of these individual mechanisms to the organ-level function of the lungs are not well understood. We developed a novel biophysical model using a compressible visco-hyperelastic formulation that incorporates pulmonary surfactant dynamics and quantifies the contributions of surfactant and parenchymal tissue viscoelasticity to lung compliance. The effect of pulmonary surfactant was modeled as a surface energy function, and the surface behavior was coupled to the bulk behavior, assuming uniform spherical alveoli. The model was used to simulate respiration and investigate the effects of altered surface tension due to surfactant dysfunction, as well as varied viscous behavior. The model captured the characteristic sigmoidal inspiratory pressure-volume relationship and hysteresis expected of lung inflation mechanics. In addition, it predicted physiologically consistent changes associated with surfactant dysfunction and alterations in tissue viscoelasticity during lung injury. We expect this work to serve as a step toward deconvoluting and predicting the respective contributions of the lung parenchyma and pulmonary surfactant to global and regional lung compliance in health and disease.
